基於深度學習發展自動車道置中控制應用於多車交通情況之自主駕駛

Abstract

近年來，自駕車已展現出在道路安全方面帶來重大改進的潛力。同時，許多基於人工智慧的自動駕駛技術被提出，用於從人類數據中學習駕駛任務。然而，針對複雜交通情況下的無人車，要達到人類水平的可靠性和安全反應是一項挑戰。 本文提出了一種自動車道對中系統的深度學習系統，該系統能夠處理多車互動場景。為了避免學習良好駕駛策略的障礙，尤其是在現有端到端方法中使用有限的專家駕駛數據的情況下，我們的系統將自動駕駛控制分為速度和轉向規劃器。此外，為了應對由於高度動態的交通場景和道路用戶交互而造成的複雜性，本論文使用強化學習架構來訓練這兩個規劃器，即使從其真實環境中收集到的數據有限，也可以有效地改善自動駕駛策略。本研究主要目標為，開發的自動車道居中系統可以通過練習新收集的數據和更新駕駛技術表示來模仿駕駛員的行為，從而提高其性能。為此，本研究使用CarSim車輛模擬軟體以及Python進行協同模擬，用於從人類駕駛員模型中學習複雜的駕駛技能的過程。實驗結果驗證了該方法在多車輛交通場景中的良好性能。實驗表明，在具有不同車輛和路況的不同軌道上，車道置中控制具有穩定而準確的性能。In recent years, autonomous vehicles have exhibited the potential to bring major improvements in road safety. Meanwhile, a range of autonomous driving technologies based on artificial intelligence are presented for learning the driving tasks from human data. However, designing for autonomous vehicles under complex traffic situations reveals challenging to reach human-level reliability and safe reaction. This thesis proposes a deep learning framework for the automatic lane centering system, which is able to handle multi-vehicle interactive scenarios. In order to avoid the obstacles for learning a good driving policy especially with limited expert driving data in existing end-to-end methods, our system breaks the autonomous driving control into a speed planner and a steering planner. Further, to confront the complexity due to highly dynamic traffic scenarios and road user interaction, a reinforcement learning framework is utilized to train these two planners so that the autonomous driving policy can be efficiently improved even with a limited collected data from the its real environment. In our main goal, the developed automatic lane centering system can improve its performance by imitating the driver behavior through practicing and updating the driving skill representation using the newly gathered data. To this end, this study builds a co-simulation platform between CarSim and Python for learning process of sophisticated driving skills from human driver models. The experimental results validate the promising performance of the proposed approach in multi-vehicle traffic scenarios. The conducted experiments with the comparison analysis of the learned system also show the stable and accurate generalization in lane centering control across various tracks with different vehicles and road conditions.