汽车工程 ›› 2021, Vol. 43 ›› Issue (8): 1195-1202.doi: 10.19562/j.chinasae.qcgc.2021.08.010

• • 上一篇    下一篇

基于RetinaNet及优化损失函数的夜间车辆检测方法

张炳力1,2,秦浩然1,2,江尚1,2(),郑杰禹1,2,吴正海3   

  1. 1.合肥工业大学 汽车与交通工程学院,合肥 230041
    2.安徽省智能汽车工程实验室,合肥 230009
    3.浙江百康光学股份有限公司,嘉兴 314113
  • 收稿日期:2021-04-07 修回日期:2021-05-19 出版日期:2021-08-25 发布日期:2021-08-20
  • 通讯作者: 江尚 E-mail:shang.jiang@hfut.edu.cn
  • 基金资助:
    安徽省第五批“特支计划”、合肥工业大学智能制造技术研究院科技成果转化及产业化平台建设专项资金(2019);科技成果转化及产业化重点项目(2019);安徽省新能源汽车产业创新发展和推广应用政策支持研发创新项目(皖发改产业函(2020)477号);学术新人提升计划A项目(JZ2021HGTA0162);青年教师科研创新启动专项A项目(JZ2021HGQA0237)

A Method of Vehicle Detection at Night Based on RetinaNet and Optimized Loss Functions

Bingli Zhang1,2,Haoran Qin1,2,Shang Jiang1,2(),Jieyu Zheng1,2,Zhenghai Wu3   

  1. 1.School of Automobile and Traffic Engineering,Hefei University of Technology,Hefei 230041
    2.Anhui Engineering Laboratory of Intelligent Automobile,Hefei 230009
    3.Zhejiang Bicom Optics Co. ,Ltd. ,Jiaxing 314113
  • Received:2021-04-07 Revised:2021-05-19 Online:2021-08-25 Published:2021-08-20
  • Contact: Shang Jiang E-mail:shang.jiang@hfut.edu.cn

摘要:

为解决智能驾驶系统中夜间车辆检测误检多、远处小目标检测效果差的问题,在RetinaNet的基础上对损失函数进行全面优化。在分类损失函数方面,分析了负样本与正样本交并比的产生机理和对训练的影响,构造了关联交并比的分类损失函数,利用负样本交并比使网络注重于训练难分类负样本,同时利用正样本交并比提高了检测框的定位精度;在定位损失函数方面,改进了传统L1损失的归一化方式,提高了小目标检测能力。此外,针对夜间场景中的车辆特征对网络结构进行了优化设计,并在夜间车辆数据集上进行了测试验证,结果表明模型优化后的平均检测精度提升了14.6%。

关键词: 夜间车辆检测, 损失函数, 交并比, RetinaNet, 小目标检测

Abstract:

In view of the high false detection rate and poor detection results for distant small target in vehicle detection at night for intelligent driving system, the loss functions are comprehensively optimized based on RetinaNet. For classification loss function, both the generation mechanism and the influence on training results of the IoUs of negative and positive samples are analyzed, the classification loss function of correlated IoU is constructed, the IoU of negative sample is used to make network lay emphasis on the training of negative samples, which are hard to classify, meanwhile the IoU of positive sample is used to enhance the locating accuracy of detection frame. For locating loss function, the normalization way of traditional L1 loss is improved and the ability of small target detection is enhanced. In addition, a design optimization is conducted on network structure for the vehicle features in night scene, and a test verification is performed on vehicle data at night. The results show that the average detection accuracy of the model optimized increases by 14.6 percentage point.

Key words: vehicle detection at night, loss functions, IoU, RetinaNet, small target detection