动态条件模仿学习自动驾驶算法研究  

Research on Automatic Driving Algorithm Based onDynamic Condition Imitation Learning

作　　者：张兴波 石朝侠[1] 王燕清 ZHANG Xingbo;SHI Chaoxia;WANG Yanqing(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094;Institute of Information Engineering,Nanjing Xiaozhuang University,Nanjing 211171)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094 [2]南京晓庄学院信息工程学院,南京211171

出　　处：《计算机与数字工程》2023年第3期545-548,598,共5页Computer & Digital Engineering

基　　金：国防科技创新特区火花课题(编号:2016300TS009091);国家自然科学基金面上项目(编号:61371040)资助。

摘　　要：针对自动驾驶领域当前条件模仿学习模型在动态环境下表现较差的问题,提出了一种使用LSTM网络融合历史视觉信息的动态条件模仿学习模型DCIL(Dynamic Conditional Imitation Learning)。首先DCIL通过残差网络Resnet34进行环境感知得到图像特征,再将连续的4帧历史图像特征通过LSTM网络进行特征融合,得到动态环境的特征向量。针对不同的导航条件,使用不同的决策网络预测车辆速度和方向盘角度;最后使用比例积分控制模型控制车速。在自动驾驶仿真平台CARLA上的测试结果表明DCIL与经典模仿学习方法相比,自动驾驶水平大幅提高,动态环境下避障能力显著增强。In order to solve the problem of poor performance of conditional imitation learning model in the field of automatic driving in dynamic environment,a dynamic conditional imitation learning model DCIL is proposed,which uses LSTM network to fuse historical visual information.Firstly,DCIL obtains image features through residual network Resnet34,and then fuses four con⁃secutive historical image features through LSTM network to obtain the feature vector of dynamic environment.According to different navigation conditions,different decision networks are used to predict vehicle speed and steering wheel angle.Finally,proportional integral control model is used to control vehicle speed.The test results on the auto driving simulation platform CARLA show that DCIL can greatly improve the level of automatic driving and the ability of obstacle avoidance in dynamic environment.

关 键 词：自动驾驶 LSTM 模仿学习 神经网络 比例积分控制 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]