We know the secret of your success
There is so much excitement amongst consumers globally, as a result of the emergence of autonomous cars for public usage. These cars do not require any human intervention or control for them to operate. According to Campbell et al. (2010), autonomous cars can identify navigation paths, sense and interpret sensory information from their local environment, classify different objects that are detected as well as the obeying of transportation rules. Great strides have been made in responding to unforeseen circumstances where an external environment does not behave as anticipated by the internal prototype or in response to backlashes that occurs in the vehicular systems. Successful autonomous navigation is made possible in these kinds of situations by combining various technologies from different fields such as control engineering, electrical engineering, electronics engineering, mechanical engineering, computer science among others (Desphpande, 2014).
The origin of autonomous cars can be traced back to 1926 with the first-ever radio-controlled car, called 'Linriccan Wonder' (Bimbraw, 2015). After this, there have been considerable improvements in autonomous car technology following the emergence of the vision-guided Mercedes-Benz robotic van in 1980 (Bimbraw, 2015; Sheng et al., 2013). Vision-guided systems that utilise LIDAR (Light Detection and Ranging) (Levinson et al., 2011), computer vision, GPS and radar, have been the area of concentration from this time, i.e. 1980. The autonomous technologies that are present today, such as steer assist, lane parking, adaptive cruise control, among others are developed from this same idea.
Different automobile organisations have forecasted that one day, completely autonomous cars will become part of everyday life. Globally, accidents from transportation are among the leading causes of death. When there are newer and innovative systems and investments in road safety at all levels, then 5 million human casualties, and 50 million severe injuries could have been prevented by 2020. The Commission for Global Road Safety believes that the number of horrendous increases in road injuries and avoidable accidents needs to be stopped and a plan put in place to reduce it yearly (Campbell, 2010). Deshpande et al. provided an estimate of approximately 3,000 deaths caused by road accidents daily with over half of the victims not travelling inside the car. Furthermore, Deshpande (2014), also stated that there would be an increase to 2.4 million victims annually if there are no significant and practical steps taken to tackle this issue, which will make road accident the 5th most common cause of death globally. Therefore, the increased reliability and speedier reaction time in comparison to human by autonomous cars will result in a dramatic decrease in the number of traffic collisions.
Additionally, there will also be a drastic reduction in traffic jam. It will increase the capacity of the roadway because there will be a lesser need for problems related to safety, and traffic flow would be managed better. It will help resolve the issue of shortage of parking since cars can drop off passengers and then park at suitable places, and later pick up these passengers which will help reduce the number of parking space needed.
It will reduce the need for visible road signage since the autonomous cars will get the relevant information through a network. The reliance on traffic police will also reduce, and as such, government expenditure will not be necessary for hiring traffic police with the use of autonomous cars. It will also help minimise the need for vehicle insurance, as well as a reduction in car thefts. It can help implement effective transportation of goods and car-sharing systems (for example, trucks and taxis respectively), entirely removing passenger redundancy. It will help relieve people from going through the stress of driving and navigating, especially people that do not enjoy driving. Moreover, it decreases commute time, since the speed at which autonomous cars travel is higher with minimal errors. An autonomous vehicle's occupants will enjoy a smoother ride in comparison to that of a non-autonomous vehicle.
At the moment, several investigators are conducting researches on autonomous driving both from academia as well as commercial projects in involving Tesla, Uber and Google self-driving cars. Autonomous cars are among the most expensive cars, even though they are designed for the comfort of human beings.
2.0 Aim and Objectives
This project aims to study the technical approaches adopted in implementing autonomous driving. The objectives involved are discussed below;
The research questions involved in performing secondary research are discussed below:
In order for highly autonomous vehicles (HAVs) to develop, the release of a guideline for autonomous driving was provided by the National Highway Traffic Safety Authority (U.S. Department of Transportation, 2018). The guideline contained the six levels of automation that SAE International stipulated (Lin et al., 2018).
In summary, the automation of both level 1 and 2 are mostly referred to as driving assistance because they still involve a considerable number of human driver intervention in a significant part of the driving tasks at all conditions. From level 3-5 of the automation, autonomous driving systems handle the entire driving responsibility in certain driving conditions; this is usually known as HAVs.
Autonomous driving has remarkably improved within the last ten years. So far, there are two main paradigms for the systems of autonomous driving, including
Mediated perception method
In the mediated perception method (Ullman, 1980), several many sub-components, are useful for the recognition of objects that are relevant to driving including pedestrians, cars, traffic lights, traffic signs, lanes among others (Geiger et al., 2013). Afterwards, the outcomes of the recognition are combined into a reliable global representation of the immediate environments of the vehicle (Figure 1).
Figure 1: Paradigm for autonomous driving (Chen et al., 2015)
All the information will be considered by the A.I. engine to control the vehicle before deciding on anything. The comprehension of the entire scene can include unnecessary complexity to a task that is already difficult because most of the parts of the objects detected are relevant to the driving decisions (Xu et al., 2017). Autonomous driving differs from other robotic cars, in that the only necessary thing for it to do is to manipulate the speed and the direction (Chen et al., 2015). The dimension of the last output space is extremely low. Mediated perception, on the other hand, processes a high-dimensional world representation, that could include irrelevant information. It will be better to directly predict the distance to a car, rather than detecting a car’s bounding box before estimating the distance to the car via the bounding box.
Moreover, computer vision perceives the individual sub-tasks included in mediated perception as open research questions. Mediated perception may include the present ultramodern methods regarding autonomous driving. It is necessary for most of these systems to depend on laser rangefinders, radar, GPS, as well as an extremely accurate map of the surroundings, so that objects can be reliable in a scene. The system is more expensive and complicated when it needs to resolve the easier problems related to the controlling of the car (Chen et al., 2015).
Behaviour reflex methods
Behaviour reflex methods use the sensory input of a driving action to create a direct mapping. It is not a new idea (Pomerleau, 1989; Pomerleau, 2012) and neural networks use it to create a direct mapping of an image to steering angles. The model’s learning process requires a human driver to drive the car along the road so that the system can record the steering angles and images as training data. The idea may have difficulty in handling traffic and complex driving manoeuvres for so many reasons, even though it is brilliant. The first reason is that other human drivers driving on the road may make different decisions, even on occasions whereby input images seem to be identical. The consequence of this is an ill-posed that will not be clear in the process of training a regressor (Chen et al., 2015).
For example, in order to overtake a car that is directly ahead of you, the driver can decide to overtake using the right or the left side (Chen et al., 2015). The presence of these situations in the training data makes it difficult for the machine learning model to make the right decision when faced with almost identical images. The second reason is that behaviour reflex has very low-level decision-making (Chatterjee, and Matsuno, 2001). It is impossible for direct mapping to have a broader view of the circumstance. For instance, the perception of the model regarding overtaking a car, and returning to a lane is simply the structure of extremely low-level decisions for manoeuvring the steering slightly to one direction, then in the opposite direction for a particular time-frame (Chen et al., 2015). This abstraction level does not adequately portray what is happening, which increases the task difficulty (Fernando et al., 2017). Lastly, the learning algorithm must decide the parts of the image that is necessary since the input image is the entire image. Nevertheless, the supervision level of training a behaviour reflex mode, that is, the steering angle, may not be strong enough to ensure that the algorithm learns this vital information (Chen et al., 2015).
The status of the current autonomous driving industry was performed by Lin et al. (2017) research was carried out by surveying the leaders of the industry on automation level, their leveraged sensors, and the computing platform. The levels achievable by even the top players of the autonomous driving organisation, such as Waymo and Tesla, are only still able to achieve level 2 or 3 of automation which still require the intervention of a human driver, to a large extent. What this implies is that building autonomous driving cars is difficult. Hence, the reason for studying this developing application. When considering the sensors and computing platforms utilised by the management of these industries, it is evident that majority of them combine both GPUs and SoCs to make a large number of computational ability that autonomous driving systems require available.
Furthermore, it was observed that the sensing devices used by Waymo, and Nvidia/Audi, which can build autonomous driving cars with level 3 automation for experiments is Light Detection and Ranging (LIDAR) driven. It is a remote sensing device that can analyse the car’s environment at high accuracy; it does so by sending light beams. LIDAR is perceived to be an excellent sensing device used in autonomous driving systems as a result of its high precision. However, the primary reason for its lack of usage in commercial purpose is because it is costly. LIDAR devices that are used commercially are sold at USD 75,000 (Lin et al. 2017), and it is a lot costlier than the cost of the car itself, or even some luxury vehicles.
Consequently, LIDAR devices are avoided by the industry but instead, they focus on building autonomous driving systems that are vision-based; that utilises only radars and cameras. Such environmental sensing devices cheap. An example can be found in organisations such as Tesla (Tesla.com, 2018) and Mobileye (Mobileye, 2018), that recently announced their intention to build autonomous driving systems that are based on vision, and that detects environments using primarily cameras and radars.
There is a general perception that vision-based solutions have potentials in resolving the problems associated with autonomous driving because the sensing cost is low. Its potentials also include the fact that computer vision is undergoing a series of development processes. Different approaches have been studied regarding the range of vision-based driving model, beginning with end-to-end approaches to full pipeline approaches (Janai et al., 2017). From the first time it was successfully demonstrated in the 1980s (Thorpe et al., 1989), the first work was done in end-to-end learning for autonomous driving, and it first emerged in 1989 and is called the ALVINN system (Pomerleau, 1989). What it indicates is that steering on simple road conditions is possible using an end-to-end model. The small fully-connected ALVINN is the network architecture developed into convolutional networks that DAVE system utilises (Yang et al., 2018), as well as the deep models that DAVE-2 system utilises (Bojarski et al., 2016). Performance can be improved using intermediate representations such as attention maps and semantic segmentation masks (Xu et al., 2017).
The control of the car and the scene parsing is separated using the pipeline approaches. The function of (Chen et al., 2015), is first to train a vehicle detector so that it can identify the location of nearby cars and use simple control logic to output the commands of the car. Huval et al. (2015) indicate that it is possible to use convolutional neural networks to detect real-time lane and cars. Even though these types of approaches make it easier to interpret, and controller, other the other hand, intermediate representation’s annotation is costly.
Figure 2: D.U. drive model architecture (Yang et al., 2018)
The transformation of the real input image to a fake virtual image that the predictor network P uses to predict the command of the vehicle is done using the generator network G. Efforts are made by the discriminator network D to differentiate between the authentic virtual images and the fake virtual images. The generator G is driven by the prediction objective as well as the adversarial object to generate the virtual representation, which produces the best prediction performance. The activation layers and instance normalisation are excluded, once all the convolutional/fully connected layers are in place, for easiness. (Abbr: k: kernel size, n: number of filters, s: stride size) (Yang et al., 2018).
Sun et al., (2018) suggests that there should be a unification structure between a real and a virtual domain for autonomous driving or DU-Drive, which transfers real driving images to their more accessible forms, in the virtual domain, using a conditional generative adversarial network. By so doing, the prediction of the control of the vehicle can be made. It is possible to train a real-to-virtual generator independently for every real domain in situations using multiple real datasets. It is also possible to train a global predictor by simultaneously using multiple sources. The outcomes of the quantitative and qualitative experiment reveal that this model can unify real images from various sources and transform them into more efficient representations. It can also get rid of the domain shift, and improve on control command prediction task’s prediction.
Figure 3: Hierarchical structure (Sun et al., 2018)
A summary of the hierarchical structure, as well as the recommended fast integrated planning and control structure for autonomous is illustrated in Figure 3. The structure includes perception, decision-making and control. The perception module senses the nearby surroundings in all the steps. It also results in the estimates/measurements of all the relevant conditions of the ego-car. For example, it is possible to forecast or detect the velocity, orientation, the location as well as the positions and velocities associated with all the other visible road participants (moving or static). According to the pre-defined driving tasks and all perceived information, the reference land and the horizon objective respectively will be determined by the decision-making module, in order to train the next-level planning and control (Sun et al., 2018)
The hierarchical framework is also used by the planning and control module, that this research recommends. The policy layer’s first layer comprises the extraction of the feature, as well as a neural network. According to driving decisions, the irrelevant and complex perception output is transformed into an abstract driving situation which a group of highly-representative features can successfully explain, using the feature extraction. After that, the features are inputted into the neural network so that it can speedily produce reference routes that emulate the optimal routes that the long-term MPC provides. Lastly, the execution layer receives such reference routes that produce resultant control actions. It achieves that by resolving the problem associated with the optimisation of a short-horizon so that it can ensure the safety and feasibility in the short-term. The recommended planning and control structure can respond quite speedily, and at the same time, achieve a comparable long-term smoothness and safety with the professional MPC policy, because of its learning-based feature (Sun et al., 2018)
It is possible to emulate the optimal driving policy that long-term MPC provided by choosing a set of features and customised Dagger process that are highly-represented. It can arrange for secured movements very quickly with long-term smoothness. Situations such as straight-going, car-following and overtaking were verified. Besides, the introduction of the virtual features can further complicate the learned policy’s driving conditions. The discussions on the virtual adjacent-lane and virtual curbs vehicles for car-following and multiple-lane driving are included in this report. Concerning future work, there will be an extension of the discussions on circumstances such as curvy roads, and intersections. Furthermore, it is possible to use the learned policy on multiple autonomous cars (a multi-agent system), as well as how to study the relationship between the agents (Sun et al., 2018).
In summary, the technology and approach used for autonomous driving have increased in the last decades. The approach for autonomous driving can be categorised into two major categories which are mediated perception methods and behaviour reflex methods (Chen et al., 2015). Even though there is technology advancement, there are no fully reliable autonomous driving cars at an affordable price for regular usage.
Bimbraw, K., 2015, July. Autonomous cars: Past, present and future a review of the developments in the last century, the present scenario and the expected future of autonomous vehicle technology. In Informatics in Control, Automation and Robotics (ICINCO), 2015 12th International Conference on (Vol. 1, pp. 191-198). IEEE.
Bojarski, M., Del Testa, D., Dworakowski, D., Firner, B., Flepp, B., Goyal, P., Jackel, L.D., Monfort, M., Muller, U., Zhang, J. and Zhang, X., 2016. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316.
Campbell, M., Egerstedt, M., How, J.P. and Murray, R.M., 2010. Autonomous driving in urban environments: approaches, lessons and challenges. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 368(1928), pp.4649-4672.
Chatterjee, R. and Matsuno, F., 2001. Use of single side reflex for autonomous navigation of mobile robots in unknown environments. Robotics and Autonomous Systems, 35(2), pp.77-96.
Chen, C., Seff, A., Kornhauser, A. and Xiao, J., 2015. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE International Conference on Computer Vision (pp. 2722-2730).
Deshpande, P., 2014. Road safety and accident prevention in India: a review. International journal of advanced engineering technology, 5(2), pp.64-68.
Fernando, T., Denman, S., Sridharan, S. and Fookes, C., 2017, October. Going deeper: Autonomous steering with neural memory networks. In Computer Vision Workshop (ICCVW), 2017 IEEE International Conference on (pp. 214-221). IEEE.
Geiger, A., Lenz, P., Stiller, C. and Urtasun, R., 2013. Vision meets robotics: The KITTI dataset. The International Journal of Robotics Research, 32(11), pp.1231-1237.
Huval, B., Wang, T., Tandon, S., Kiske, J., Song, W., Pazhayampallil, J., Andriluka, M., Rajpurkar, P., Migimatsu, T., Cheng-Yue, R. and Mujica, F., 2015. An empirical evaluation of deep learning on highway driving. arXiv preprint arXiv:1504.01716.
Janai, J., Güney, F., Behl, A. and Geiger, A., 2017. Computer vision for autonomous vehicles: Problems, datasets and state-of-the-art. arXiv preprint arXiv:1704.05519.
Levinson, J., Askeland, J., Becker, J., Dolson, J., Held, D., Kammel, S., Kolter, J.Z., Langer, D., Pink, O., Pratt, V. and Sokolsky, M., 2011, June. Towards fully autonomous driving: Systems and algorithms. In Intelligent Vehicles Symposium (IV), 2011 IEEE (pp. 163-168). IEEE.
Lin, S.C., Zhang, Y., Hsu, C.H., Skach, M., Haque, M.E., Tang, L. and Mars, J., 2018, March. The Architectural Implications of Autonomous Driving: Constraints and Acceleration. In Proceedings of the Twenty-Third International Conference on Architectural Support for Programming Languages and Operating Systems (pp. 751-766). ACM.
Mobileye. (2018). Mobileye | Autonomous Driving & ADAS (Advanced Driver Assistance Systems). [online] Available at: https://www.mobileye.com/ [Accessed 14 Dec. 2018].
Pomerleau, D.A., 1989. Alvinn: An autonomous land vehicle in a neural network. In Advances in neural information processing systems (pp. 305-313).
Pomerleau, D.A., 2012. Neural network perception for mobile robot guidance (Vol. 239). Springer Science & Business Media.
Sheng, W., Ou, Y., Tran, D., Tadesse, E., Liu, M. and Yan, G., 2013, November. An integrated manual and autonomous driving framework based on driver drowsiness detection. In Intelligent Robots and Systems (IROS), 2013 IEEE/RSJ International Conference on (pp. 4376-4381). IEEE.
Sun, L., Peng, C., Zhan, W. and Tomizuka, M., 2018, September. A fast integrated planning and control framework for autonomous driving via imitation learning. In ASME 2018 Dynamic Systems and Control Conference (pp. V003T37A012-V003T37A012). American Society of Mechanical Engineers.
Tesla.com. (2018). Electric Cars, Solar Panels & Clean Energy Storage | Tesla. [online] Available at: https://www.tesla.com/ [Accessed 14 Dec. 2018].
Thorpe, C., Hebert, M., Kanade, T. and Shafer, S., 1989. Vision and Navigation for the Carnegie Mellon Navlab. In High Precision Navigation (pp. 97-122). Springer, Berlin, Heidelberg.
Ullman, S., 1980. Against direct perception. Behavioral and Brain Sciences, 3(3), pp.373-381.
U.S. Department of Transportation. (2018). USDOT Automated Vehicles Activities. [online] Available at: https://www.transportation.gov/AV [Accessed 13 Dec. 2018].
Xu, H., Gao, Y., Yu, F. and Darrell, T., 2017. End-to-end learning of driving models from large-scale video datasets. arXiv preprint.
Yang, L., Liang, X., Wang, T. and Xing, E., 2018, September. Real-to-Virtual Domain Unification for End-to-End Autonomous Driving. In Proceedings of the European Conference on Computer Vision (ECCV) (pp. 530-545).
Last updated: Mar 29, 2020 07:43 AM
Your one-stop website for academic resources, tutoring, writing, editing, study abroad application, cv writing & proofreading needs.
