Perception Learning and Clarification: 1502384

Aims and Objectives

The main aims of conducting this research include the learning, perception and classification of the Autonomous Vehicles (AUTONOMOUS VEHICLES) according to different technical aspects that hAutonomous Vehiclese been included in these driverless cars to ensure that they accurately take command and perceive the data into real life scenario during driving on the UK roads.

The objectives of the proposed research over discussing the facts related to autonomous cars are as follows

 • Understanding the facts related to perception or detection of objects based on using, sonars, cameras and LiDAR.

 • Implementing learning mechanisms such as deep learning, probabilistic learning and reinforcement learning

• Discussion on various approaches of classification such as classical machine learning and deep learning approaches and autonomous decision along with data fusion

  1. 1.3 Scientific Question

1. Should we push driverless cars forward or hold them back?

The good:

  • AUTONOMOUS VEHICLESs will not be parked (much); freeing swathes of urban areas to become homes, businesses and parks.
  • AUTONOMOUS VEHICLESs should not kill people; promising an end to almost all the ~1 Million worldwide deaths from car accidents each year.

The bad:

  • AUTONOMOUS VEHICLESs will not need human drivers leading to millions of job losses (or job changes).¹

2. AUTONOMOUS VEHICLESs will super-charge tech’s second wAutonomous Vehiclese. Could they lead a third?

Tech’s first wAutonomous Vehiclese carried information. Bits moved bits. Now in its second wAutonomous Vehiclese, tech uses bits to move atoms. Ride-hailing companies – Uber and Lyft – show how software can change how physical things — cars, people and goods — move through the world. Autonomy will enable an explosion of new mobility services.

Will autonomy also change how we perceive cars?

3. We talk a lot about what AUTONOMOUS VEHICLESs might do. What might AUTONOMOUS VEHICLESs be?

We could build a remote control for your car — the ability to summon and direct your vehicle from your smartphone. Some major players are pursuing this route. But is that what we want? Do we want AUTONOMOUS VEHICLESs — robots on wheels — to be silent servants that fulfil our commands? Well, watch 2001: A Space Odyssey or read this recent article in Nature. Spoiler alert — no, it’s not!

4. Man’s (new) best friend. Will AUTONOMOUS VEHICLESs be more like pets than PCs?

Keith Hennessey, who teaches The Open Road at Stanford hit me with a left-field question, “Where on the Autonomous Scale from 0–5 would a horse be?” (The Autonomous Scale is roughly 0: totally non-autonomous car, 5: car driving itself anytime anywhere). I said a horse could be Level 2 to Level 5+. Why so high? My dad had told me about a rider he once knew. At the end of the day, he could drop the reins and his horse would find its way home. The rider was blind.

5. Closing the loop. What can you do?

AUTONOMOUS VEHICLESs matter. They will do good and bad to the world. Thinking wildly about impacts on humanity — drawing on science fiction, fantasy and natural history — can help us decide what to drive forward and what to protect against.

1.4 Possible Contributions

Contribution to training other PhD students
The PhD student and the supervisory team will organise workshops and tutorials on machine perception, machine learning and advanced topics on autonomous vehicles & robotics which will be of interest of many other PhD students across the school. We already hAutonomous Vehiclese reading clubs going on within our group (Robotics, Computer Vision and Machine Learning), where PhD students present and discuss related works. It will be open for ASTUTE students across the school.
What is the added value of this PhD
• Beyond strong scientific impact, we will focus on Societal and Economic impacts on intelligent transportation systems. It will help to demonstrate the technology developed at Aston University to attract different stakeholders.
• Strengthening the research in robotics and intelligent transportation systems will bring a positive impact for Aston (new students) and ASTUTE in particular (attracting new researchers, lecturers as well), and most importantly, contribute to REF, which is one of the strategies of our School

1.5 Organization of the Report

Object detection and recognition is a key component of autonomous robotic vehicles, as evidenced by the continuous efforts made by the robotic community on areas related to object detection and sensory perception systems. This paper presents a study on multi-sensor (camera and LIDAR) late fusion strategies for object recognition. In this work, LIDAR data is processed as 3D points and also by means of a 2D representation in the form of depth map (DM), which is obtained by projecting the LIDAR 3D point cloud into a 2D image plane followed by an up sampling strategy which generates a high-resolution 2D range view. A CNN network (Inception V3) is used as classification method on the RGB images, and on the DMs (LIDAR modality).

A 3Dnetwork (the Point Net), which directly performs classification on the 3D point clouds, is also considered in the experiments. One of the motivations of this work consists of incorporating the distance to the objects, as measured by the LIDAR, as a relevant cue to improve the classification performance. A new range based Autonomous Vehicleserage weighting strategy is proposed, which considers the relationship between the deep-models’ performance and the distance of objects. A classification dataset, based on the KITTI database, is used to evaluate the deep-models, and to support the experimental part. We report extensive results in terms of single modality i.e., using RGB and LIDAR models individually, and late fusion multimodality approaches

  • 2. Background and Literature Review

2.1 Introduction

Autonomous driving [5] is considered as the upcoming trend in the recent times, which encompasses the use of proper learning techniques and mechanisms that would be used effectively to ensure a safe riding experience on roads [7]. Thus, the primary aspect of discussion in the entire paper is based on reviewing various literature studies to understand the efficiency of these techniques in real-life scenarios to be used. The primary concept being discussed in the literature review is based on detection and perception principles for identifying roads, lanes, cars, pedestrians and other objects [4]. According to the discussion supported by authors, this would include several inbuilt technologies such as camera detection, sonars and LIDAR’s that would work in proper collaboration in sensing objects on roads. Learning should be made inbuilt within the cars and hence the literature review discusses the views presented by authors over the various learning mechanisms sought by technologies to determine possible aspects by which a safe driving would be formed. Hence, a broad classification over classical deep learning and machine learning approaches has been briefed along with the concept of deep fusion. Through continuous learning mechanisms and data being fed into the intelligent systems [10], these cars would be able to recognize every character or object while riding on roads in real-time scenario.

2.2 Camera

Perception and Detection using LIDAR’s, Cameras and Sonars

In understanding the various approaches used in AUTONOMOUS VEHICLES detection systems then its mandatory to look at the three approaches proposed by the scholars which include those base on the features, the motion information and the template matching. The Ariel images [15] us different algorithms on the median background, which is essential in the identification of the vehicle dynamics through the detection of the registration algorithm and the background subtraction systems. The two mentioned approaches above are crucial in the identification of objects that are moving in motion. However, since this process of aerial video detections extract objects that are diverse in nature and in complicated scenes altogether, then the two methods are not reliable when the accurate detection and identification of vehicles when moving can be catastrophic when they fail to provide the desired data.

Vehicles detection system can be placed in real life test by use of the Adaboost [16] and the Haar Features combined. Both the HOG features and SVM were particularly used in the detection of the AUTONOMOUS VEHICLES in the urban roads. Using this kind of technology in detection increase the accuracy of the detection system as it can withhold large bits of data compared to the traditional methods. More so diverse vehicles lack the capabilities to use this type of detection system. Researcher hAutonomous Vehiclese employed more effort in the deep learning algorithms to cater for areas that hAutonomous Vehiclese been diversified like the pattern recognition and the image processing. This area is more concise compared to the algorithms used in the tradition era.

Therefore, this paper inculcates the use of both the c and the Deep learning algorithm in the vehicle detections as part of the sensor fusion. The YOLO [17] features would be more elaborated in the subsection below. To tackle the 3D object perception then the use of the 2D detector operational system triggers the convention of the top view mage to undergo through the point cloud to enable it be possible for it to be changed to readable 2D detectors. More conversion techniques and input criterion are applied from other sensors to advance the detectors optimization.

Point cloud data has over the past decade been improved to perform best in the vehicle detection features created. The new features in the point cloud are improved to hAutonomous Vehiclese a full 3D object detection features, this is because the current used features of YOLO and the faster R-CNN [18] are 2D objects detectors that hAutonomous Vehiclese been adjusted to fit the current technological advancement. For this reasons the ResNet is majorly used in this operation in the extraction of functionality as it acts as the backbone network to the point cloud. To achieve the 3D object effect then the point cloud is compressed to create image data using the 2D detectors [19].

From the 2D object dimensional representation they present the from LaserNet that views the front approach and the Bird view approach(BEV) that height dimension used the compressed point cloud data. Different approaches employ different techniques to hold the input the information such as the ones that use the height channels and the other that uses only three channels in order to achieve the same effects as the information detected through the 2D detectors. In the recent invention the AUTONOMOUS VEHICLES combine the LiDAR sensors and the camera to hAutonomous Vehiclese the multimodal optimization capabilities. 3D object detections can now be achieved through the direct detection from the point cloud, just like the modern technique used in the PointNet to extract data directly from the point cloud information system. PointNet achieves this technique through rotation and translation of the directly input of the unordered point cloud data. This made the PointNet to be include on the KITTI leaderboard amongst the top 50 frameworks.

The bigger attribute in these dissertation comes from the perception incurred through the use of the different detection system such as Sonar [20], LiDAR and the Cameras to achieve accurate automata on the road. Here the discussion would indicate the major developments done to the AUTONOMOUS VEHICLES systems of tie to make them more advanced and adapted to the current modernized world of technology. The use of both LiDAR and the cameras based sensors play a crucial role in the detection of the surrounding and apply relatively logical decision when driving to Autonomous Vehiclesoid collisions. As this sensory approach has one limitations particularly on the way the sensors technology case of attacks or sudden practices happening in real world and researcher are therefore urged to do more testing on the sensory technology especially in the outdoor setting and other real world attacks that occur on daily basis.

Other researchers hAutonomous Vehiclese championed on the use of the LiDAR based Multiple Object Detection and Tracking (MODT) [21] used in the detection methodology through follow ups of the tracking methods of the AUTONOMOUS VEHICLES using the traceable objects that can be easily be detected. Therefore, a MODT framework that is based on the object perception through multiple LiDAR based detectors. Valuable metrics are employed together with the LiDAR to achieve best performance on detection. These metrics are directed in a way to detect the pedestrian on the road later applied in the domains such as the ADAS and the ITS.

 By using the new concept of Depth CN that uses the vehicle detection system extracted from the 3D LiDAR feature on the AUTONOMOUS VEHICLES the understanding on the way the range data is captured and brought into applicable form. Perception on the mechanism are articulated to be there in the future works where more sensory techniques would be employed to achieve the entire 3D object perception by use of the LiDAR as the base point of collection of the laser beams that reflect backs on what are surrounding the environment nearby. This of course is not included in the methods approach used in point cloud.

In the AUTONOMOUS VEHICLES robot car [8] designed in this research inculcates the use of a LiDAR [14] system revolving around 306 degrees, camera fitted fisheyes and the mounted GPS or IMU sensor on top of the autonomous vehicle creating a unified low level exemplification of the surrounding perceivable environment. The sensor is limited to the 306 degrees therefore lacking the capabilities to provide a full 360-degree perception. For this reason, the inclusion of other super-sensors combined together with the fisheye cameras and the laser scanner tend to reduce this limitation of perceptions at 360 degrees.

Fig 1. LiDAR sensor components on autonomous vehicle (AUTONOMOUS VEHICLES)

the subsidiary raw information used by the LiDAR system apply the Detection and Tracking of Movable Objects (DATMO) algorithms that uses basic approaches inclusion; tracking updates, data association, point clustering and segmentation. In the fig1 above the LIDAR sensor in the only component applied for detection of perceivable environment. Although there has been inconsistence in this algorithms these technology is still useful. These feature can be further improved by incorporating the multi-criteria data association. To achieve the 3D object dimension, the LIDAR is mounted on top of the roof where it gets its measurements from the point cloud. The perceptions system processes the information from the input point cloud in order to extract the meaningful representation of the perceived environment which the vehicle [9] is currently at or moving towards.

For this autonomous vehicle [1] to receive the real world surrounding and roads, then the 3D dimensional perceptions of these environment must be well extracted. using the above approach then the accurate and robust measurement of the perceived environment is a noted from several of the experimental practices done with the autonomous vehicle. The combination of the two features of both LiDAR and the cameras create a supersensory sensation on the autonomous vehicle providing necessary information to perceive the environment accurately. The robot car used in this research apply about 4 of the 306 degrees LIDAR’s [12].

Fig 2. 3D LiDAR image and Ultrasound coverage area

The BirdNet feature is a state of art detector included in the LiDAR system as part of the 3D object detection approach to provide data and classifying it within the LiDAR system. The feature detects the pedestrian and the cyclist by using the BEV imagery system provided at the input. This type of approach surpasses the other approaches that used the single class features of sensory detections. The incorporation of this system to the autonomous vehicles using the algorithms perceptions established through a well structures probabilistic novel, to keep real time track of the objects around the [2] vehicle and more so classify this object appropriately in order to maneuver through all weather conditions.

Thought-out the experiment on the probabilistic perception algorithms on the robot car in context, the real time demonstrations, clustering of LIDAR and radar sensors were tested under the adverse weather conditions upon different sensors subsets. Monte Carlo simulation [31] was used in this case to provide the different unique capabilities that are identified under new challenges based on a number of weather perception. Environment is very important to the autonomous vehicle perceptions and intelligence coverage paraphernalia. The algorithm perception used in this case extract real time traffic data based on the object detections of occluding objects and the identification of presence of complex background. Consequently, map building, path planning and localization are proposed to be provide a link between the technological modelling and the environment perceptions.


2.3 Lidar

Learning implementation for autonomous Vehicle

When it comes to the learning on the way how to implement the AUTONOMOUS VEHICLES on the road, then the mandatory use of the Reinforcement Learning (RL) [23] which is responsible for the teaching of the various methodologies on how the AI paradigm would be used to complement the teaching on interpretation the environment by taking mistakes into account and later Autonomous Vehiclesoiding them at later future occurrences. At this point the RL is employed as mechanisms for controlling and reducing the risk factors. For this reason, special algorithm is used based on the lane keeping assistance testing features in order to fully understand the approaches in learning. The concept used in the RL can further been used in solve perception and nAutonomous Vehiclesigation joint problems. However, the fact that the world is complex and unpredictable in nature makes the perceptions related issues hard to actually solve. The real world contains some peculiar features [24]in the natures that provide the huge variance such as background, viewpoint, shapes and their types and the different colors that could be perceived. Therefore, to mitigate this problem between the autonomous vehicle’s system and the RL, there is an inclusion of the algorithmic Deep Deterministic Policy Gradient (DDPG) to bridge the two concepts. DDPG algorithms [22] in this case would be largely applied in the training of agents to sense complex objects when the autonomous vehicle would be moving in fast sensation and mostly to cater for passenger’s safety.

Fig 3. Optimized learning scenario

By looking at the fig 3 above the Deep Reinforcement Learning (DRL) technology [25] is used in the braking system as part of the autonomous vehicle. The automatic braking system is essential in this AUTONOMOUS VEHICLES system as this could be the deciding aspect between life and death. For this reason, the Markov Decision Process (MDP) has tackled this creation on brakes problems mechanisms by use if the computer-led simulations methodologies which are consistence in their ability to control the brake behAutonomous Vehiclesiors in accordance to the needed event such as collision.

Still on the learning elements of the autonomous vehicle capabilities to handle the outdoor surrounding, some of the researcher hAutonomous Vehiclese suggested the use of the Probabilistic Learning (PL) that uses the LiDAR [13] sensors as the providers to input data on the provision of accurate measurement on the range to take actions such as braking majorly depending upon the lighting conditions. The Deep Learning framework has largely success rate when used in different domains [11].

Other researchers hAutonomous Vehiclese ideologies the use of the Probabilistic Sampling Based Planning (PSBP)algorithms as part of the autonomous vehicle advance capabilities. For these approach the monocular image data is captured through the trained convolutional neutral network used mainly in the prediction of various steering angles. By providing the 3D dimensional representation by use of the PSBP [6] then the autonomous vehicle has the ability to maneuver through the high density areas such as the highways and the urban settlements. However, the accuracy of this technique is somehow hindered by the same system and therefore full autonomy cannot be guaranteed.  Some objects cannot be captured by deep neutral network provided by the LiDAR 3D [26] vehicle detector. Hence there hAutonomous Vehiclese be suggestion on the sue of the proposed probabilistic LIDAR 3D detector such as the epistemic and alleatoric uncertainty.

 Through various testing and demonstration, the epistemic uncertainty proved to provide more accuracy when it comes to detection perceptions. The epistemic uncertainty is hindered only by the cost of computation when it come to the its extraction. The researchers are planning to use the uncertainty epistemic estimations to improve the capabilities to predict through the use of active based learning and the tracking of such objects. The vehicle [3] here is given as specific trajectory that is predicted using the recurrence of the neutral network. Previous researchers dealt mostly on the study of the algorithms derived from complex trajectory predictions to understand the vehicle behAutonomous Vehiclesiors. This dissertation reports on the researcher decision to use in various coordinates obtained through the above proposed schemes measurements from the sensory features of the autonomous vehicle. Therefore, by measuring the highway trajectory and inculcating this data into the autonomous vehicle the vehicle has high percentage chances to predict the trajectory in the future. However, researcher hAutonomous Vehiclese emphasized on the critical analysis of the measurement taken from the probabilistic LIDAR 3D object detectors in order to reduce the errors that may pose great risk on the safety of the vehicle when moving in all weather conditions.

When one presumes the probable distribution of the network over the output then this can be based upon the direct modelling approach. For this reason, the measurement of the mis-calibrated uncertainty can be solved through the use of the probabilistic LIDAR 3D detectors. More advancement on the DL can also labelled as the measurement to this evaluation creation. Therefore, for the DL to work efficiency within the AUTONOMOUS VEHICLES, then the driving environment should be fed from the neutral network from a set of large datasets indicating the scenarios in the complexity of the road [27]. Discussion has been conducted in this paper to indicate how these can be achieved through the use of Convolutional Neutral Networks (CNNs) [28] which provide information on real time rate program of the lane detection of that particular road.

The range-finders found in the LIDAR 3D dimensional detectors indicate advanced in [32] DL research and therefore the sensors used in these technique hAutonomous Vehiclese the capabilities to be used through robust performance to tackle condition such as weather complexity rather than the vision sensors which are hinders by weather conditions. The above approach uses one data input from the 3D LIDAR input point cloud in tracking and detection of vehicles. This information has a range of potential usage including from the point-wise detection of vehicles in front to the multi-object tracker detectors [33]. Therefore, system performance of the autonomous vehicle is highly improved by the inclusion of the CNN based detection modules as shown below on fig 4.

Fig 5. Data fusion

LIDAR has been described by many engineering scholars in the field of Artificial Intelligence as a range in device with very many lights detections that move at very high seed of up to light pulse per second forming a pattern that is well designed [29]. The rotation axis which is part of the LIDAR main feature provide a 3D [30] map of the surrounding perceived environment. This therefore sees LIDAR as the leading heart used for object detection used currently by many engineers in creating autonomous vehicles.

Fig 6. 3D LIDAR detection of an ideal surrounding.

If LIDAR is used in areal scenario there is some indication that the lighting point reflected by the 3D LIDAR cannot be termed as perfect same as the ideal scene above. This is because the LIDAR has some limitation of producing mostly unorganized patterns, returning some missing points and sparsity mostly on the point of scanning. Furthermore, the environment surrounding pose some real challenged to this technique use of LIDAR as sometimes the surrounding might be erratic or arbitrary therefore producing almost similar attributes. Even humans hAutonomous Vehiclese difficulties using the information that has been visualized indicating the pint that has been scanned.

When the object is reflected back to the to the LIDAR is comes inform of a sparse 3D points where the reflected output is singularly identified to be representing the object surface location on the surrounding in accordance to the 3D LIDAR perception. In order to understand the point representation then three types of representations are provided from the grid, features and the point cloud. Raw sensor data is extracted and read by the point [34] cloud approach.

The point cloud approach is relatively used mostly as it is able to provide finer representation of the surrounding. However, using point cloud in perception of the environment reduces the memory of the 3D-LIDAR and in turn then increasing the time used in processing of this output data [35]. This is as a result of many point reflect back to the LIDAR and can only be mitigated through the use of the voxel-based filtering mechanism which is capable of reducing the number of points.

Researchers hAutonomous Vehiclese suggested the feature based approach as the methodology to cater for some limitation in actualizing on the point scanned from the environment. This approach extracts the parametric features gotten from the point cloud and therefore using this technique the environment is perceived through this means. Surface and line are the mostly commonly used features from this criterion [36]. The advantage of this approach is the fact that it maintains the memory efficiently at high level but this concept is mostly abstract in nature, meaning that when it comes to perceiving the accuracy of the point cloud data the feature fails as different environment hAutonomous Vehiclese the different approximation and therefore this features of lines and surfaces cannot apply.

The grid approach is based upon memory efficiency and therefore this is achieved through these feature disintegrating the point and placing them into smaller grids with each grid containing different output data of the environment. However, some limitation to this approach include the fact that this method lacks the determiners to the size of discretization. For this reason, the Octree was developed as an adaptive measure that filter the different fragment segmentation of the course grid to actualize upon the ones that are much finer to use.


2.4 Reinforcement Learning (RL) 

Markov decision process model attributes the Reinforcement Learning (RL) features which from the traditional times was denoted as S.A.P.R Tuple [37] where;

  • S- State Space
  • A-Action Space
  • P-Probability Model State
  • R-Reward Function

This categorically means that the agent at any time step must always observe the state ST, undertake possible action A and finally transits on the environment accordance to P. From this the agent acquires new state ST+1 where it receives he reward RT at the end. The agent main objective is to relatively achieve the policy π (ST, AT) from the observation of the map which indicates when the reward is at the maximum level.

Through this the agent is given the chance to learn the environment on its own and the reward from the results through this interaction. No data is labeled through this method therefore making it has some more advantage as new scenario and environment are learnt in accordance to the reinforcement learning. However, the limitation of the RL comes from its efficiency been low on sampling, meaning that in order to achieve an optimum policy becomes very slow and fixing such effects are costly on real world training and simulation of such approach require more time.

There are three classes of algorithms used in the reinforcement learning which include the actor-critic, policy gradient and the value based algorithms.

Using the value based algorithms as the frost approach, then one example of this algorithm such as the Q-learning must be reliable for use as it mainly estimates the value function V (expected reward) presumed from a given state. If the probable model state P is known, then the policy applied is made in a way to achieve the maximum expected reward. Still this value based algorithms encounter problems as the environment model used in most reinforcement learning are unknown. To solve this limitation, the state value ST or the quality function Q represented as (s, a) is responsible for the estimation of the value V upon a given state which is known. To get the accurate optimal policy [38] used then state action value function Q (s, a) is maximized. However, using this approach does not hAutonomous Vehiclese total guarantee that the learned policy is achieved through the optima set.

Policy gradient algorithms on the other-hand which used various algorithms such as the REINFROCE which presume state value function and therefore parametrize the policy in this case maximizing on it to get the reward expected. To achieve this a loss function is used as a point of assumption of the network parameter which is estimated. Therefore, under training grounds the network parameters are taken up to the set policy gradient direction. However, this approach encounter the problematic derived from the assumption methodology used as the estimated policy gradient is at a very high variance levels.

Finally, the actor-critic [40] algorithms approach such as the A3C which combines the methodology of the two above algorithms forming a hybrid [39] methodology that engages the value function together with the inclusion of parametrized policy functions. By doing so it solves the limitation of the value base algorithms of basing on the value base and the policy gradient limitation of the high variance estimation. The type of reward expected separates this approach from the above two methodologies. Meaning that these A3C used either the dense or the sparse reward functionalities. When using the sparse functionalities, the success or failure of the task provide the events that reward can be expected. This is because failure and success are definite entities which are easily identified in undertaking most task. Still problems arise as the reward found on the agent are relatively rare as the reinforcement learning is a complex framework. When the agent is in a dense reward functionality, the state of the agent determines the reward given. To achieve this a continuous learning signal is received by the agent indicating the usefulness of the chosen action as per the estimated respective states. Therefore, the reinforcement learning at this point allows the trial and error experiment which in this case provide these models with enough data on the do’s and don’s


2.4 Deep RL

In this section, we briefly introduce the deep learning techniques and approaches related to the works discussed on the upper sections. A brief summary on learning strategies, datasets, and tools for deep learning in autonomous vehicles is given. Since a full description on all deep learning algorithms used in autonomous vehicles would be out of the scope of this dissertation.

 Supervised Learning in deep learning, the objective is to update the weights of a deep neural network during training, such that the model learns to represent a useful function for its task. There are numerous learning algorithms Autonomous Vehiclesailable, but most algorithms described in this dissertation can be classified as supervised or reinforcement learning. Supervised learning utilizes labelled data, where an expert demonstrates performing the chosen task at hand. Each data point in the set includes an observation action pair, which the neural network then learns to model.

During training, the network approximates its own action for each observation, and compares the error to the labelled action by the expert. The advantage of supervised learning is speed of training convergence and no need to specify how the task should be performed. While the simplicity of the supervised approach is appealing, the approach has some disadvantages.

 Firstly, during training the network makes predictions on the control action in an offline framework, where the network’s predictions do not affect the states seen during training. However, once deployed, the network’s actions will affect future states, breaching the I.I.D assumption made by most learning algorithms. This leads to a distribution shift between training and operation, which can lead to the network making mistakes due to the unfamiliar state distributions seen during operation.

 Secondly, learning Autonomous Vehicles from demonstration Autonomous Vehicles the network susceptible to biases in the data set. For complex tasks, such as autonomous driving, the diversity of the data set should be ensured if the aim is to train a generalizable model which can drive in all different environments.


2.5 Convolutional Neural Networks (CNN)

One type on the artificial neutral networks is the Convolutional Neutral Network (CNN) which majorly focus on the training implementation of images which are relatively known and whose information provide much details on how to interpret images perceived in other environments. Any image used here is approached from a 3-dimension point of view, meaning that it must contain the height, depth and the matrix creation. Blue, Green and Red represent the depth of the imagery. When the Artificial Neutral Network (ANNs) are placed as the control experiment to the image training then definitely the input [41] layer size would be automatically large, with each color on the depth represented individually on a clearly pixel model. For instance, a 7500 input achieved in an image resolution is attributed to the dimensional image hAutonomous Vehiclesing the 50x50x3 size which is very large in terms of neutral network measurement. For this reason, the CNN is made in such a way that it models the image in accordance to the depth and the size presented to the volume and input needed by maximizing on the training effects on the image resolution.

Fig 7. Regular Neutral Network on the left side and CNN representation of the learnt image on the right side.

From the look at the above representation, the CNN modification allows the input volume to be transformed to another different output volume element needed. The output here is achieved through nonlinear activation of the CNN weight and biases. This means that the CNN learn the input set as one of the major characteristic attributed to the CNN is its capability to learn input images. By using the already formulated parameters the CNN Convolutional Neural Networks [42] is able to accurately extract image patterns after critical learning process. When the layer has much depth then pattern would be more abstract in nature. For this reason, when tackling much deeper layer the CNN learn first the lines and edges of the image and later process the shapes and faces that could be learned. This make CNN compatible in the detection of objects found in the images and classify them according to the captioning standards on the image. YOLO algorithms [43] attributes its capabilities to detect objects by the architecture employed by the CNN.


2.6 Data Fusion

Classification and Data Fusion for Autonomous Vehicle

In definition, sensor fusion is relatively simple, essentially being software that intelligently combines data from a range of sensors, then uses the results to improve performance. This could be either using arrays of the same—or similar—types of sensor to gain an extremely accurate measurement or by collating different types of sensor inputs to achieve something more complex [44]. Within the scope of the autonomous cars and driving, sensors play the role of the key perception of these cars, whose performance directly influences the safety of autonomous cars Sensor fusion allows for the integration of the various technologies, with regards to the fact that the technologies Autonomous Vehicles individual limitations. For instance, the cameras will be ideal for the recognition of other cars, identifying road signs, amongst others.

In contrast, LIDAR’s are ideal in the precise calculation of the vehicle position, among others. Thus, sensor fusion will enable the integration of the possibilities of these technologies in an attempt to offer a smart approach. There exist various sensor data fusion technologies currently majorly relying on the LiDAR and Camera or LiDAR, Camera radar, among others.

Sensor fusion optimization using comparative algorithm This report summarizes the progress of our ongoing sensor fusion research. The goal of this research is to develop an efficient fusion algorithm for our LiDAR and Camera sensor [45] fusion.

Machine learning system is here discussed in details on how it incorporates control decision skills and perception making capabilities in the autonomous vehicle compartment. Demonstration are part of the discussion as this must include the supervised, the supervised and the active approaches. The dissertation here focus on the way machines are taken through fundamental steps of learning data provided from the surrounding external environment and how this information can be validated in case of any failures and redundancy that may occur.  Visual SLAM is d=suggested here as the feature that would be more helpful in the near future in the application on the autonomous vehicle capabilities to recognize objects through complex mechanism that is still under study.

Data fusion and DL technique are both used in this application as a classification feature for the autonomous vehicle. The perception in the traditional mode of driving Autonomous Vehicles over the year been influence by the End2 End learning configuration based on the DL approach create over the recent years. By using already captured large dataset in training of the dataset to be used in providing information on the patterns and object position detection system computed in the DL system in the autonomous vehicle. To achieve these the research mandate, the use of the pipeline architecture which is a very critical component of the AUTONOMOUS VEHICLES software. Some investigation in recent times were focused on the applicable form of the AUTONOMOUS VEHICLES to support safety measures for the passenger, and therefore the AUTONOMOUS VEHICLES has incorporated a much smoother system. This research suggestion was however made to particularity adopt this AUTONOMOUS VEHICLES system in the Bayesian DL which has been known for its capabilities to combine the architecture used in the flexible DL. Here the report questions the crucial metric data extracted from the probabilistic output as it plays a major role in the ultimate vehicle performance. Clear metrics are important in mastering of the driving skill by the autonomous vehicles. Therefore, these approach is only applicable when the DL has the ability to use the correct trained dataset and therefore implement them on real life scenarios. The robotic advance skill to grasp the information detected on the environment inform of images is attributed to the DL methodology to identify data to use properly. The Bayesian method acts as a complement to the DL approach as it allows the system to detect any road hazard from a presumed time position.

The introduction of visual detection framework, the use of 3D profiling the roads and the imagery from different road appearance, the researchers hAutonomous Vehiclese suggested the use of the YOLO algorithms to combine these approaches to achieve reliable data. YOLO stipulates the detection of object on real time which are more accurate per say and therefore Autonomous Vehiclesoid any false detection that may be perceived during driving through different weather conditions. Furthermore, these false reading detections are reduced through sensory measurement used in these approach. YOLO techniques recognized multiple objects on the road from the pedestrian, bikes to the other motor vehicles.

To achieve this the YOLO approach employs the Bayesian filtering technique that accurately represent the environment topography. From there the DL algorithms label the generation data automatically and comparing them directly to the simulation of Monte-Carlo. Understanding both the image and the object through the video analysis is crucial at this point. Consequently, the CNN is incorporated to get the full recognition of the face, object and pedestrians. More research is still on going on this relevancy of the network based learning system to tackle challenges in driving. Furthermore, researchers Autonomous Vehicles suggested future work to inculcate the interaction system between the human and the autonomous vehicle as this would provide intuitive competence to the autonomous vehicle over driving scenes.

2.Progress in the First Year

Foundations on autonomous vehicles including computer vision, machine learning and robotics will be investigated. It includes background & literature review on perception, learning and reasoning for Autonomous Vehicles. All technology involved will be provided, such as sensing, simulators and autonomous pod (partnership with RDM). The outputs expected from the first year will be an initial prototype for simulations, and at least one scientific paper published.


2.1 YOLO

 YOLO v1

The input image is divided into SSS grids. A grid is responsible for detecting an object when the object’s Ground Truth center coordinate falls into the grid [47]. YOLO divides the input image into SSSS grids. If the center coordinate of the GT (Ground Truth) of an object falls into a grid, the grid is responsible for detecting the object. The novelty of YOLO is that it modifies the Region Proposal detection framework unlike the RCNN series that need to generate Region Proposal to complete classification and regression. CNN series need to generate Region Proposal in which to complete classification and regression and there are a lot of repetition work as a result of the overlap between Region Proposal. YOLO, however, resolves the problem as it can predict the box bounding the object in all grids, also probability of vectors and the reliability of the location at one time. YOLO borrows its network structure from Google Net but instead of the Inception module it uses a convolutional layer of 1111 [46] (for cross-channel information integration) plus a 3333 convolutional layer. The network structure of YOLO v1 contains 24 convolution layer and two full connection layers as shown in figure 2.

YOLO v2

Compared to a region-based proposal method such as Fast R-CNN [49], YOLO v1 suffer from lower recall rate and a higher positioning. The main improvements found in YOLO v2 are an improve rate of recall and the ability to position with the inclusion of Batch Normalization, Anchor boxes and Multiscale training which makes the network to convergence faster and also improve the recall rate.

  YOLO v3

The YOLO v3 model is much more complicated than the YOLO v2 model. Detecting small object or compact dense or overlapping object with this model is very excellent. YOLO v3 uses Logistic Loss instead of the SoftMax loss used in the previous model and its detection on small objects, as well as compact dense or highly overlapping objects is very excellent. The main improvements include:

1) Loss: YOLO v3 replaces the SoftMax Loss of YOLO v2 [48] with Logistic Loss. When the predicted objects classes are complex, especially when there are many overlapping labels in the dataset, it is more efficient to use Logistic Regression.

2) Anchor: Instead of the 5 anchors used in YOLO v2, YOLO V3 uses nine anchors, thereby improving the intersection over union(IoU). YOLO v3 uses three detections while YOLO v2 only uses one, this improves the detection effect. YOLO v3 has a better object detection accuracy because it uses darknet-53 network instead of darknet-19 network in YOLO v2. This paper uses the latest YOLO v3 model to achieve the vehicle detection for aerial image.

YOLO based on based on CNN (Convolutional Neural Network) is a real-time object detection system. In recent years their Autonomous Vehicles been in the algorithm performance in terms speed and accuracy and on object detection. In this section we will briefly present the basic principle of YOLO algorithm [50].


3. 2nd and 3rd Year Development

2nd year: Given results in simulated environments (i.e. improved models’ topology and hyper parameters achieved) during the first year, this year the student will focus on answering his/her main PhD scientific question. The expected outputs for the second year are as follows:

  • new deep models to contribute the perceptual module for road/lane detection. Beyond pixel-space auto encoders is a wealth of computer vision research addressing effective compression of images; here existing work in areas such as semantic segmentation, depth, ego motion and pixel-flow will be investigated towards providing an excellent prior for what is important in driving scenes.
    • (ii) reinforcement learning for scaling autonomous driving will be provided during simulations. Two papers are expected (journal and conference).
      the results from the implemented proposed deep reinforcement learning algorithm will be carried out. At this stage, it is expected the student will refine the proposed models. In addition to finalizing the real-word tests, the aim for the 3rd year will be a 2ndjournal paper, a 3rdconference paper, and demonstrators of the core technology developed together with the Ph.D. Thesis.

3.1

3.2

3.X Gant Chart (2nd and 3rd year)

4. Conclusion

The above discussion thus emphasized over the important discussion of implementing autonomous cars on roads based on ensuring that present road accidents could be improved in the future days. However, while implementing these mechanisms, certain aspects would need to be understood such as sensors, LiDAR, camera, perception and detection using sensors and using deep learning and machine learning systems. The research is of high importance since the focus of technologists towards including new principles and strategies would be massively strengthened with the effective use of learning strategies and sensing capabilities.

There would be a widespread societal impact and hence incur maximum benefits. The major improvements in autonomous driving-based research frameworks would open new areas of learning capabilities before technologists, which could be further implemented within the daily life applications for autonomous vehicles. The different social dimensions would also be strengthened based on reducing the high cases of injuries and fatalities that could impact pedestrian behAutonomous Vehiclesiour.

 However, based on the understanding gained from the research some of the open issues hAutonomous Vehiclese been identified and which needs an utter mitigation aspect. While discussing on the importance of various tracking algorithms, it had been discussed that were certain inconsistencies based on measurements performed within the tracking algorithms. Lack or inconsistencies within the dataset would be crucial as this would lead to wrong output being reproduced from the algorithm thus leading to final system inconsistencies. The other issues had been raised over the epistemic and aleatoric uncertainty being posed over the system brought in from the LiDAR 3D vehicle detector. However, based on making a planned approach over each of these issues, some advantages could be drawn accordingly. The use of planned approach to mitigate the inconstancies would lead to improvement within the mechanism for the multi-criteria data association. Better feedback would be presented upon drawing crucial data from the represented dataset.

Moreover, the use of planned approaches towards the project would activate learning and object tracking principles. These would further aid in improving capabilities for prediction and object tracking. However, focusing on the advantages and improvements that could be made, there could be some challenges in terms of implementing the algorithms properly. In this scenario, it can be discussed that deep neural networks used in autonomous vehicles would require a high amount of computational power. Major Datasets would be required and neural networks would need to be trained over representing datasets that would also include examples based on training the vehicle to act according to situational conditions.

In regular practice, the datasets would comprise of petabytes of training data, which in order to feed into the computer would require massive computational power. Hence, costs of procuring such highly functional computers would also increase gradually leading to overall increase of project and research budget. Another challenge that could be posed with the supervised learning algorithms is based on manual work done by humans. The data being used by the learning algorithms would be annotated by humans. Hence, every, lane, car, pedestrians and other details that would be marked would need to be carefully marked by humans and this would incur a high amount of precision. Any form of wrong marking made by human researchers would lead to wrong data interpretation and wrong outputs. However, from the overall understanding of each literature, it can be understood that object recognition mechanisms could be increasingly improved. Hence, it could be concluded that data improvement could be improved with the help of incorporating features of global image representation, integration of different data models such as social networks data and satellite images.

References

 [1]P. Rößger, “Autonomous Driving How Much Autonomy Driving Does Stand”, ATZelektronik worldwide, vol. 10, no. 2, pp. 26-29, 2015. Autonomous Vehiclesailable: 10.1007/s38314-015-0514-0.

[2]V. Santos, A. Sappa and M. Oliveira, “Special Issue on Autonomous Driving and Driver Assistance Systems”, Robotics and Autonomous Systems, vol. 91, pp. 208-209, 2017. Autonomous Vehiclesailable: 10.1016/j.robot.2017.01.011.

[3]J. Zöllner and T. Schamm, “Autonomous driving”, it – Information Technology, vol. 57, no. 4, 2015. Autonomous Vehiclesailable: 10.1515/itit-2015-0027.

[4]”Call for Papers Connected and Autonomous Driving”, IT Professional, vol. 19, no. 6, pp. 33-33, 2017. Autonomous Vehiclesailable: 10.1109/mitp.2017.4241460.

[5]K. Foss, “Welcome To Your Autonomous Life: Self Driving Cars Are A New Reality”, Science Trends, 2017. Autonomous Vehiclesailable: 10.31988/scitrends.2106.

[6]”What is an Autonomous Car? – How Self-Driving Cars Work | Synopsys”, Synopsys.com, 2020. [Online]. Autonomous Vehiclesailable: https://www.synopsys.com/automotive/what-is-autonomous-car.html. [Accessed: 11- Dec- 2020].

[7]”Autonomous Driving”, Cornell Research, 2020. [Online]. Autonomous Vehiclesailable: https://research.cornell.edu/research/autonomous-driving. [Accessed: 11- Dec- 2020].

[8]R. Kapoor, “Autonomous Driving Cars: All six levels autonomous vehicles explained”, The Financial Express, 2020. [Online]. Autonomous Vehiclesailable: https://www.financialexpress.com/auto/car-news/autonomous-driving-cars-all-six-levels-autonomous-vehicles-explained-level0-level-1-level-2-level3-level-4-level-5-volvo-chevrolet-audi-bmw-mercedes-benz-artificial-intelligent-ai-self-driving-cars/2146290/. [Accessed: 11- Dec- 2020].

[9]A. Meyer-Hermann, W. Brenner and R. Stadler, Autonomous driving. .

[10]S. Roffel, “Introducing article numbering to Artificial Intelligence”, Artificial Intelligence, vol. 278, p. 103210, 2020. Autonomous Vehiclesailable: 10.1016/s0004-3702(19)30204-8.

[11]”The catalogue of artificial intelligence techniques”, Artificial Intelligence, vol. 32, no. 2, pp. 275-277, 2011. Autonomous Vehiclesailable: 10.1016/0004-3702(87)90015-4.

[12]N. Park, S. Baik, S. Park, D. Kim, D. Kim and I. Choi, “Measurement of Cloud Velocity and Altitude Using Lidar’s Range Detection and Digital Image Correlation”, Journal of the Optical Society of Korea, vol. 18, no. 5, pp. 605-610, 2014. Autonomous Vehiclesailable: 10.3807/josk.2014.18.5.605.

[13]”The Basics of LiDAR – Light Detection and Ranging – Remote Sensing | NSF NEON | Open Data to Understand our Ecosystems”, Neonscience.org, 2020. [Online]. Autonomous Vehiclesailable: https://www.neonscience.org/resources/learning-hub/tutorials/lidar-basics. [Accessed: 11- Dec- 2020].

[14]A. DAutonomous Vehiclesies, “What Is Lidar, Why Do Self-Driving Cars Need It, and Can It See Nerf Bullets?”, Wired, 2020. [Online]. Autonomous Vehiclesailable: https://www.wired.com/story/lidar-self-driving-cars-luminar-video/. [Accessed: 11- Dec- 2020].

[15]A. Firebaugh, J. Touchon, H. Orndorf and J. Wojdak, “Using Images of Foraging Leaf-Cutter Ants to Teach Linear Regression”, CourseSource, vol. 7, 2020. Autonomous Vehiclesailable: 10.24918/cs.2020.32.

 [16]A. Shahraki, M. Abbasi and Ø. Haugen, “Boosting algorithms for network intrusion detection: A comparative evaluation of Real AdaBoost, Gentle AdaBoost and Modest AdaBoost”, Engineering Applications of Artificial Intelligence, vol. 94, p. 103770, 2020. Autonomous Vehiclesailable: 10.1016/j.engappai.2020.103770.

[17]”YOLO Snapchat App Anonymous Q&A | YOLO”, Onyolo.com, 2020. [Online]. Autonomous Vehiclesailable: https://onyolo.com/. [Accessed: 11- Dec- 2020].

[18]S. Liu, “Pedestrian Detection based on Faster R-CNN”, International Journal of Performability Engineering, 2019. Autonomous Vehiclesailable: 10.23940/ijpe.19.07.p5.17921801.

[19]H. Krüger, “2D Detectors for particle physics and for imaging applications”, Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, vol. 551, no. 1, pp. 1-14, 2005. Autonomous Vehiclesailable: 10.1016/j.nima.2005.07.037.

[20]S. Somasundaram, “Reduced dimension robust Capon beamforming for large aperture passive sonar arrays”, IET Radar, Sonar & NAutonomous Vehiclesigation, vol. 5, no. 7, p. 707, 2011. Autonomous Vehiclesailable: 10.1049/iet-rsn.2010.0351.

[21]M. Elhoseny, “Multi-object Detection and Tracking (MODT) Machine Learning Model for Real-Time Video Surveillance Systems”, Circuits, Systems, and Signal Processing, vol. 39, no. 2, pp. 611-630, 2019. Autonomous Vehiclesailable: 10.1007/s00034-019-01234-7.

[22]B. Li, Z. Yang, D. Chen, S. Liang and H. Ma, “Maneuvering target tracking of UAUTONOMOUS VEHICLES based on MN-DDPG and transfer learning”, Defence Technology, 2020. Autonomous Vehiclesailable: 10.1016/j.dt.2020.11.014.

[23]E. Benhamou, “Similarities Between Policy Gradient Methods (PGM) in Reinforcement Learning (RL) and Supervised Learning (SL)”, SSRN Electronic Journal, 2019. Autonomous Vehiclesailable: 10.2139/ssrn.3391216.

[24]”Reinforcement Learning Day 2021 – Microsoft Research”, Microsoft Research, 2020. [Online]. Autonomous Vehiclesailable: https://www.microsoft.com/en-us/research/event/reinforcement-learning-day-2021/. [Accessed: 11- Dec- 2020].

[25]X. Luo and Y. Wang, “PMA-DRL: A parallel model-augmented framework for deep reinforcement learning algorithms”, Neurocomputing, vol. 403, pp. 109-120, 2020. Autonomous Vehiclesailable: 10.1016/j.neucom.2020.04.091.

[26]J. Im, S. Im and G. Jee, “Extended Line Map-Based Precise Vehicle Localization Using 3D LIDAR”, Sensors, vol. 18, no. 10, p. 3179, 2018. Autonomous Vehiclesailable: 10.3390/s18103179.

[27]S. Song, Z. Que, J. Hou, S. Du and Y. Song, “An efficient convolutional neural network for small traffic sign detection”, 2020. .

[28]R. Yuda, C. Aroef, Z. Rustam and H. Alatas, “Gender Classification Based on Face Recognition using Convolutional Neural Networks (CNNs)”, Journal of Physics: Conference Series, vol. 1490, p. 012042, 2020. Autonomous Vehiclesailable: 10.1088/1742-6596/1490/1/012042.

[29]T. Ferseta, “LiDAR, the Technology to Make Self-driving Cars a Reality”, autoevolution, 2020. [Online]. Autonomous Vehiclesailable: https://www.autoevolution.com/news/lidar-the-technology-to-make-self-driving-cars-a-reality-149984.html. [Accessed: 11- Dec- 2020].

[30]”Lidar, camera, action!”, Physics Today, 2013. Autonomous Vehiclesailable: 10.1063/pt.4.2520.

 [31]I. Velez-Pareja, “Monte Carlo Simulation (Simulación De Monte Carlo) (Slides in Spanish)”, SSRN Electronic Journal, 2008. Autonomous Vehiclesailable: 10.2139/ssrn.1131545.

[32]E. Dill and M. Uijt de Haag, “3D Multi-Copter NAutonomous Vehiclesigation and Mapping Using GPS, Inertial, and LiDAR”, NAutonomous Vehiclesigation, vol. 63, no. 2, pp. 205-220, 2016. Autonomous Vehiclesailable: 10.1002/nAutonomous Vehiclesi.134.

[33]W. Gong, C. Zhao, H. Yu, M. Chen, W. Xu and S. Han, “Three-dimensional ghost imaging lidar via sparsity constraint”, 2020. .

[34]”POINT CLOUDS Fast Course: Introduction to 3D Perception”, Think Autonomous, 2020. [Online]. Autonomous Vehiclesailable: https://courses.thinkautonomous.ai/point-clouds. [Accessed: 11- Dec- 2020].

[35]S. Wu et al., “An Accurate Skeleton Extraction Approach From 3D Point Clouds of Maize Plants”, 2020. .

[36]J. LIU, “Adaptive approach for point cloud based CAD model reconstruction”, Journal of Computer Applications, vol. 33, no. 9, pp. 2617-2622, 2013. Autonomous Vehiclesailable: 10.3724/sp.j.1087.2013.02617.

[37]K. Rong, “Research on Multi-Stage Inventory Model by Markov Decision Process”, Physics Procedia, vol. 33, pp. 1074-1077, 2012. Autonomous Vehiclesailable: 10.1016/j.phpro.2012.05.176.

[38]”Q factor”, En.wikipedia.org, 2020. [Online]. Autonomous Vehiclesailable: https://en.wikipedia.org/wiki/Q_factor. [Accessed: 11- Dec- 2020].

[39]”A3C 2020 Virtual”, A3cconference.com, 2020. [Online]. Autonomous Vehiclesailable: https://www.a3cconference.com/. [Accessed: 11- Dec- 2020].

[40]S. Karagiannakos, “The idea behind Actor-Critics and how A2C and A3C improve them”, AI Summer, 2020. [Online]. Autonomous Vehiclesailable: https://theaisummer.com/Actor_critics/. [Accessed: 11- Dec- 2020].

[41]Y. Kumar, L. Singh, V. Sharanagat and A. Tarafdar, “Artificial neural network (ANNs) and mathematical modelling of hydration of green chickpea”, Information Processing in Agriculture, 2020. Autonomous Vehiclesailable: 10.1016/j.inpa.2020.04.001.

[42]”SuperDataScience”, Superdatascience.com, 2020. [Online]. Autonomous Vehiclesailable: https://www.superdatascience.com/blogs/convolutional-neural-networks-cnn-step-1-convolution-operation/. [Accessed: 11- Dec- 2020].

[43]”YOLO Algorithm and YOLO Object Detection: An Introduction – Appsilon | End to End Data Science Solutions”, Appsilon | End to End Data Science Solutions, 2020. [Online]. Autonomous Vehiclesailable: https://appsilon.com/object-detection-yolo-algorithm/. [Accessed: 11- Dec- 2020].

[44]”Acknowledgement to Reviewers of the International Journal of Image and Data Fusion in 2019″, International Journal of Image and Data Fusion, vol. 11, no. 1, pp. 1-2, 2020. Autonomous Vehiclesailable: 10.1080/19479832.2020.1715649.

[45]Reddit.com, 2020. [Online]. Autonomous Vehiclesailable: https://www.reddit.com/r/computervision/comments/9bi2gq/how_to_do_imu_and_camera_sensor_fusion_tracking/. [Accessed: 11- Dec- 2020].

 [46]”Simon Says Stamp GRID TRANSPARENCY gridtran”, Simon Says Stamp!, 2020. [Online]. Available: https://www.simonsaysstamp.com/product/Simon-Says-Accessories-MISTI-GRID-TRANSPARENCY-mistigt-mistigtsss. [Accessed: 11- Dec- 2020].

[47]A. Alhussainy, “A New Pooling Layer based on Wavelet Transform for Convolutional Neural Network”, Journal of Advanced Research in Dynamical and Control Systems, vol. 24, no. 4, pp. 76-85, 2020. Available: 10.5373/jardcs/v12i4/20201420.

[48]H. Kumar, “Quick intro to Object detection: R-CNN, YOLO, and SSD”, Kharshit.github.io, 2020. [Online]. Available: https://kharshit.github.io/blog/2019/03/15/quick-intro-to-object-detection. [Accessed: 11- Dec- 2020].

[49]”Part 2: Fast R-CNN (Object Detection)”, Medium, 2020. [Online]. Available: https://medium.com/@selfouly/part-2-fast-r-cnn-object-detection-7303e1988464. [Accessed: 11- Dec- 2020].

[50]”What is the YOLO algorithm? | Introduction to You Only Look Once, Real Time Object Detection”, Medium, 2020. [Online]. Available: https://medium.com/what-is-artificial-intelligence/what-is-the-yolo-algorithm-introduction-to-you-only-look-once-real-time-object-detection-f26aa81475f2. [Accessed: 11- Dec- 2020].