基于Dim env-YOLO算法的昏暗场景车辆多目标检测  

Multi-Target Detection of Vehicles in Dim Scenes Based on Dim env-YOLO Algorithm

作　　者：郭克友[1] 王苏东 李雪[1] 张沫[2] GUO Keyou;WANG Sudong;LI Xue;ZHANG Mo(School of Artificial Intelligence,Beijing Technology and Business University,Beijing 100048,China;Research Institute of Highway Ministry of Transport,Beijing 100088,China)

机构地区：[1]北京工商大学人工智能学院,北京100048 [2]交通运输部公路科学研究院,北京100088

出　　处：《计算机工程》2023年第3期312-320,共9页Computer Engineering

基　　金：交通运输行业重点科技项目“面向自动驾驶汽车封闭测试环境的可遥控仿真行人测试装备研发”(2018-ZD1-010);北京工商大学2021年研究生科研能力提升计划项目。

摘　　要：低照度的夜间路况复杂,现有夜间车辆识别相关研究较少,且存在识别方法实时性不高、过多占用硬件资源等不足。针对夜间场景车辆识别干扰因素较多、检测效果不佳的问题,提出一种基于YOLOv4的Dim env-YOLO车辆目标检测算法。利用MobileNetV3网络替换原始YOLOv4中的主干网络,以减少模型参数量。在改进的YOLOv4模型上使用图像暗光增强方法,提高车辆目标在昏暗环境中的可识别性。在此基础上,引入注意力机制加强特征信息选择,同时利用深度可分离卷积降低网络计算量。选取北京部分道路的夜间场景图片自制数据集并进行实验验证,结果表明,在存在高斯噪声、模糊扰动、雨雾夜晚等情况下,Dim env-YOLO算法的测试结果较稳定,对于照度低于30 lx的昏暗条件下的车流,其检测mAP值达到90.49%,对于最常见的轿车类别,mAP值达到96%以上,优于Faster-RCNN、YOLOv3、YOLOv4等网络模型在昏暗光照条件下的检测效果。Nighttime road scenes with low illuminance are complex,and visibility is typically poor.However,relatively few works have considered vehicle recognition under these conditions,and existing methods require considerable computational resources and cannot operate in real time.To address these problems,a Dim env-YOLO vehicle target detection algorithm based on YOLOv4 is proposed.Use MobileNetV3 network to replace the backbone network in the original YOLOv4 to reduce the amount of model parameters.The improved YOLOv4 model performs low-light image enhancement to improve the identifiability of vehicle targets in dark environments.An attention mechanism is implemented to enhance the selection of feature information,and depthwise separable convolution is used to reduce computational resource requirements.A dataset of images of nighttime road scenes in Beijing was collected to experimentally evaluate the proposed approach.The results show that the performance of the Dim env-YOLO algorithm was relatively stable under conditions of Gaussian noise,fuzzy disturbance,rain,fog,and so forth,and it achieved an mAP value of up to 90.49%in recognizing vehicle flow under dim condition with illumination of less than 30 lx.For the most common category representing cars,the mAP value of the proposed method exceeded 96%,which outperformed conventional methods including Faster-RCNN,YOLOv3,YOLOv4,and other network models.

关 键 词：昏暗场景 车辆检测 深度可分离卷积 Dim env-YOLO算法 MobileNetV3网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]