无人驾驶中运用DQN进行障碍物分类的避障方法  

Obstacle Avoidance Method Using DQN to Classify Obstacles in Unmanned Driving

作　　者：刘航博 马礼[1] 李阳 马东超[1] 傅颖勋 LIU Hangbo;MA Li;LI Yang;MA Dongchao;FU Yingxun(School of Information,North China University of Technology,Beijing 100144,China)

机构地区：[1]北方工业大学信息学院,北京100144

出　　处：《计算机工程》2024年第11期380-389,共10页Computer Engineering

基　　金：国家自然科学基金(62272007,62001007);北京市自然科学基金(4234083,4212018)。

摘　　要：安全是无人驾驶汽车需要考虑的首要因素,而避障问题是解决驾驶安全最有效的手段。基于学习的避障方法因其能够从环境中学习并直接从感知中做出决策的能力而受到研究者的关注。深度Q网络(DQN)作为一种流行的强化学习方法,在无人驾驶避障领域取得了很大的进展,但这些方法未考虑障碍物类型对避障策略的影响。基于对障碍物的准确分类提出一种Classification Security DQN(CSDQN)的车辆行驶决策框架。根据障碍物的不同类型以及环境信息给出具有更高安全性的无人驾驶决策,达到提高无人驾驶安全性的目的。首先对检测到的障碍物根据障碍物的安全性等级进行分类,然后根据不同类型障碍物提出安全评估函数,利用位置的不确定性和基于距离的安全度量来评估安全性,接着CSDQN决策框架利用障碍物类型、相对位置信息以及安全评估函数进行不断迭代优化获得最终模型。仿真结果表明,与先进的深度强化学习进行比较,在多种障碍物的情况下,采用CSDQN方法相较于DQN和SDQN方法分别提升了43.9%和4.2%的安全性,以及17.8%和3.7%的稳定性。Safety is the primary factor to consider for unmanned driving vehicles,and obstacle avoidance is the most effective means to ensure driving safety.Learning-based obstacle avoidance methods have attracted the attention of researchers owing to their ability to learn from the environment and make decisions directly from perception.Deep Q-Network(DQN)has made significant progress as a popular reinforcement learning method in obstacle avoidance for autonomous driving;however,it does not consider the impact of the obstacle category on obstacle avoidance strategies.Therefore,we propose a vehicle-driving decision-making framework,the Classification Security DQN(CSDQN),which is based on accurate obstacle classification results.This framework aims to achieve higher safety in autonomous driving by finding safer decision strategies based on different obstacle types and environmental information.First,the detected obstacles are classified according to their safety levels,and then safety evaluation functions are proposed for different types of obstacles.The uncertainties in position-and distance-based safety measures are used to evaluate safety.The CSDQN decision-making framework utilizes obstacle categories,relative position information,and safety evaluation functions for continuous iterative optimization to obtain a final model.Finally,the proposed method is compared with advanced deep reinforcement learning.The simulation results show that in the presence of multiple obstacles,the CSDQN method improves safety by 43.9%and 4.2%and stability by 17.8%and 3.7%,respectively,compared to the DQN and SDQN methods.

关 键 词：无人驾驶 深度Q网络 分类避障 评估函数 安全性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]