A brief overview of some noteworthy exploited Machine Learning techniques to address the self-driving challenges.
This post is the second part of a series of articles. The first part can be accessed from the "related articles" section below.
Planning
Planning refers to a set of techniques designed to make decisions to move a vehicle from a starting point to a predefined destination. It can be divided into three separate tasks, namely High-Level (or Global) Planning, Behavioral Planning and Path (or Local) Planning.
Global Planning defines the routes and the travel itinerary to follow. Typically, it is performed searching the shortest point-to-point path in a graph representation of the road network. According to the desired requirements the edges of the graphs, interpretable as road segments, can be differently weighted modifying the outcomes.
Dijkstra’s and A* are two valuable Graph Search algorithms adopted in a wide range of applications for decades. Although they work well with small graphs, for larger problems more complex methods including Deep Reinforcement Learning are often employed.
Behavioural Planning means generating local objectives by executing two complementary sub-tasks, i.e. prediction and decision-making. In prediction, the outputs of the perception phase are utilised to compute the trajectory and the future position of an obstacle, e.g. a pedestrian.
Some popular deep learning approaches are DeepSort and ROLO. Both of them track objects by associating the entities detected in different frames. Whereas the former exploits a pre-trained neural network along with Kalman filters and the Hungarian algorithm, the latter adds Recurrent Neural Network (RNN) units to the YOLO architecture.
On the other hand, decision-making ensures the car respects traffic laws and interacts appropriately with the other agents. Although Finite State Machines are oftentimes used in this context, they are limited because manually designed for a finite set of possible situations. Deep RL is consequently emerging as a preferred and more powerful solution.
Local Planning aims to generate a smooth trajectory to accomplish the local objectives defined by Behavioural Planning while avoiding obstacles and collisions. The algorithms employed, such as Rapidly-exploring Random Trees (RRT), Probabilistic Roadmaps (PRM) and RL, are usually evaluated and compared on the bases of their:
completeness, i.e. whether the algorithm terminates in finite time;
computational efficiency, namely the execution time and scalability of the method.
Control
In summary, Perception, Localisation and Planning provide respectively the knowledge of the environment, the position of the vehicle in it and the trajectory to follow to avoid collisions and respect laws. The translation of the planned intentions to actual movements is hence performed by the Control module.
A Control algorithm, alternatively known as controller, is designed to produce accurate steering angles and acceleration values to follow as precisely as possible the generated trajectory. A controller must always be able to correct the dynamic of the vehicle in case of imprecise movements or external disturbances such as rainfall, wind and slippery roads. Overall, as well as in Planning and more and more frequently in end-to-end systems, Deep Reinforcement Learning is considered one of the most promising and valuable methodologies to explore.
References
Cohen, Jeremy. "Deep Learning in Self-Driving Cars - Becoming Human: Artificial Intelligence Magazine." Medium, 12 Mar. 2021, becominghuman.ai/deep-learning-algorithms-in-self-driving-cars-14b13a895068.
Cohen, Jeremy. "Computer Vision for Tracking - Towards Data Science." Medium, 14 Oct. 2020, towardsdatascience.com/computer-vision-for-tracking-8220759eee85.
Cohen, Jeremy. "Computer Vision Applications in Self-Driving Cars - Becoming Human: Artificial Intelligence Magazine." Medium, 14 Oct. 2020, becominghuman.ai/computer-vision-applications-in-self-driving-cars-610561e14118.
Martin, Lance. "Perception for Automated Trucking - Ike Blog." Medium, 17 Aug. 2020, medium.com/ike-blog/perception-for-automated-trucking-c8a8c12e1015.
Wojke, Nicolai, Alex Bewley, and Dietrich Paulus. "Simple online and realtime tracking with a deep association metric." 2017 IEEE international conference on image processing (ICIP). IEEE, 2017.
Ning, Guanghan, et al. "Spatially supervised recurrent convolutional neural networks for visual object tracking." 2017 IEEE International Symposium on Circuits and Systems (ISCAS). IEEE, 2017.
The images in the blog are either copyright free or designed from scratch. Some figures presented in this article are also created leveraging elements extracted from the following vector images:
"https://www.freepik.com/vectors/car" Car Vector created by macrovector.
Comments