基于改进YOLOv3的车辆前方动态多目标检测算法  

Dynamic multiple object detection algorithm for vehicle forward based on improved YOLOv3

作　　者：金立生 郭柏苍 王芳荣[3] 石健 JIN Li-sheng;GUO Bai-cang;WANG Fang-rong;SHI Jian(School of Vehicle and Energy,Yanshan University,Qinhuangdao 066004,China;Hebei Key Laboratory of Special Delivery Equipment,Yanshan University,Qinhuangdao 066004,China;College of Communication Engineering,Jilin University,Changchun 130022,China;School of Mechanical Engineering,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]燕山大学车辆与能源学院,河北秦皇岛066004 [2]燕山大学河北省特种运载装备重点实验室,河北秦皇岛066004 [3]吉林大学通信工程学院,长春130022 [4]北京理工大学机械与车辆学院,北京100081

出　　处：《吉林大学学报（工学版）》2021年第4期1427-1436,共10页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(52072333);河北省重点研发计划项目(21340801D,20310801D).

摘　　要：现阶段的环境感知目标检测技术多为单类目标检测,或是将一幅图像中所有目标均列为待检目标,较少有对处于车辆前方的目标进行针对性的划分和检测。为了解决以上问题,提出了将车辆前方的待检目标分为两类:一是危险性较大,随时可能发生位移的动态目标,包括四轮车辆、二轮车辆和人;二是危险性较小,不会发生位移的静态目标,包括交通信号灯和交通标识。针对危险性较大的车辆前方动态多目标,提出了一种可以移植于嵌入式端的改进YOLOv3的目标检测算法,针对原始YOLOv3算法得到模型较大,难以在嵌入式端实时检测的缺点,以轻量型骨干网络MobileNetV2替换YOLOv3原始骨干网络Darknet-53进行特征提取,在训练中加入群组归一化操作,并使用Adam作为优化器。使用提取后的BDD100K数据集进行训练,利用未参与训练的BDD100K部分数据集和自采标注的Team_test数据集进行测试。研究结果表明,相比于原始YOLOv3算法,本文算法的漏检率可以维持在5%以内,在mAP提升0.020的基础上,本文模型在参数量上较YOLOv3基础模型减小了约89%,在CPU下的Inference Time缩小了约70%。The task of object detection plays an important role in the safe driving of driverless vehicles.Currently,the object detection technology of environment percept is mostly one-class object detection or all the objects in an image are listed as the target to be detected.Numerous studies have not yet focused on object division and detection of the objects in front of the vehicle.To solve the above problems,in this paper,the objects to be detected in front of vehicles are divided into two categories.One is the dynamic targets with high risk and displacement at any time,including four-wheel vehicle,two-wheel vehicle and people.The other one is the static targets with less danger and no displacement,including traffic lights and traffic signs.For the dynamic multiple objects in front of the vehicle,an improved algorithm of object detection based on YOLOv3 is proposed,which can be transplanted to the embedded system.To overcome the shortcoming of the original YOLOv3 algorithm,that it is difficult to get real-time detection in the embedded terminal,the original backbone network Darknet-53 was replaced with MobileNetV2 to extract features,adding Group Normalization operation in the training process and using Adam as optimizer.The extracted BDD100K dataset is used for training.The model is tested with BDD100k partial dataset not involved in training and Team_test dataset produced by our research group.The results show that compared with original YOLOv3,the missing rate(MR)of the algorithm in this paper can be kept within 5%,and based on the increase of 0.020 in mAP,comparing with the basic model of YOLOv3,the parameters of YOLOv3-MobileNetV2 model are reduced by about 89%,the Inference Time is reduced by about 70%under the CPU.

关 键 词：无人驾驶技术 环境感知 深度学习 多目标检测 轻量化模型 

分 类 号：U491.2[交通运输工程—交通运输规划与管理]