Automated lane change decision making for autonomous vehicles using machine learning techniques

Abstract

Autonomous cars play a significant role in the future of transportation. Due to the progress in Artificial Intelligence, it is anticipated that future smart cars and trucks will be driverless, accident avoiding, and efficient. They eliminate human driving errors and supply safe travel by reducing reaction time lag and safe lane change. To reach these goals, automakers have invested in the research areas regarding the current challenges to reach the expected results. Following the recent advancement in machine-learning algorithms, it is expected that driver-less cars will appear in the market within the next two decades. Self-driving cars received considerable attention during the past ten years. The concurrency of this attention with the rise of deep learning and deep reinforcement learning algorithms is not a coincidence. Deep learning algorithms found their path into the autonomous vehicle applications first by applying convolutional neural networks (CNN) to image classifications. The obtained promising results have motivated researchers in the area of the autonomous vehicles to utilize deep neural networks in the perception layer of the advanced driver-assistance systems (ADAS). The perception layer in ADAS is responsible for detecting and classifying actors surrounding the ego vehicle, e.g., cars, pedestrians, and cyclists. After creating a bird-eye-view (BEV) mapping of the environment, the sensor fusion layer identifies and tracks the other actors in the scene. Then, a decision-making algorithm produces high-level actions - e.g., ChangeLane, Accelerate, or Stop - to minimize a pre-defined cost function, usually to avoid collisions and achieve a goal location as soon as possible. Trivially, after producing optimal high-level actions, lower-level controllers generate the desired throttle and steering angle to follow the given commands. This thesis investigates autonomous lane-change by utilizing different methods such as Q-table, reinforcement learning, and neural network. The autonomous lane change is one of the crucial parts of ADAS's decision-making center, especially while driving on highways. Currently, the autonomous vehicles, that actively drive in cities cannot perform safe and reliable lane changes without relying on the human driver. Most of the companies utilize the Finite State Machine approaches to generate heuristic decisions based on the driving situation. However, these techniques may not generalize enough to capture different driving situations affected by the environment, road, and traffic conditions. Recently, Reinforcement Learning algorithms have shown promising results in producing discrete actions by observing discrete and continuous environments. Thus, we have decided to apply these algorithms to tactical decision making in highway driving tasks for self-driving cars. To this end, we have simulated the grid-world of the car on PyGame environment with four number of lanes and three different actions such as turning left, turning right, and stay on the lane. Finally, we compared the results of different methods and came up with different future work scenarios to see the feasible impact.