基于深度学习的应急装备物资车辆检测跟踪算法研究

Study on Emergency Vehicle Detection and Tracking Algorithm Based on Deep Learning

为了实现公路交通突发事件应急处置中对应急装备物资车辆的实时检测跟踪,从而提升应急装备物资车辆的调度管理效率,保障公路交通安全畅通,利用道路视频图像,研究应急装备物资车辆轨迹特征检测识别算法,并开发检测软件。首先采集来源于青海省公路交通应急装备物资储备中心辐射区域多个路段的监控、航拍拍摄的应急装备物资车辆行驶场景的视频图像序列,隔帧抽取出5293张图像作为应急装备物资车辆图像样本库,并制作目标检测的数据集;其次采用YOLOv3算法检测运动的应急装备物资车辆,通过增加网络的检测层,改进网络结构,提高模型对于应急装备物资车辆小目标的检测性能,使用SGD随机梯度优化算法训练改进后的YOLOv3算法,训练后的模型平均精度达到了95%,利用该模型实现了运动的应急装备物资车辆的精准检测;然后,结合Deep Sort算法对应急装备物资车辆进行跟踪,在图像上画出跟踪的应急装备物资车辆的运动轨迹,并对车辆进行自动跟踪计数,在多场景、多路段进行算法验证,实现了5%以内的平均误差;最后,基于提出的YOLOv3+Deep Sort结合的车辆检测跟踪算法,开发基于视频的应急装备物资车辆检测与跟踪软件系统,可实现统计车辆数和车辆的轨迹提取,并利用道路视频测试,验证了系统的检测跟踪效果。

In order to realize the real-time detection and tracking of emergency vehicles in the emergency disposal of highway traffic emergencies,and to improve the efficiency of emergency vehicle scheduling management and ensure the safety and smoothness of highway traffic,the detection and recognition algorithm of emergency vehicle trajectory characteristics is studied by using road video images,and the detection software is developed.First,the video image sequences of surveillance and aerial photography of driving scene of emergency vehicles from multiple road sections in the radiation area of Qinghai transport emergency equipment and materials reserve center are collected,5293 images every other frame are extracted as the image sample library of emergency vehicles,and the target detection data set is made.Second,the moving emergency vehicles are detected by using the YOLOv3 algorithm.By adding the detection layer of the network and improving network structure,the detection performance of the model for small targets of emergency vehicles is improved.The improved YOLOv3 algorithm is trained by using the SGD random gradient optimization algorithm,the average accuracy of the trained model reaches 95%,and the model is used to realize the accurate detection of moving emergency vehicles.Then,the emergency vehicles are tracked by using the Deep Sort algorithm,the trajectories of the tracked emergency vehicles are drawn on the images,the vehicles are automatically tracked and counted,and the algorithm is been verified in multiple scenes and multiple sections,achieving an average error within 5%.Finally,based on the proposed vehicle detection and tracking algorithm which combined with YOLOv3+Deep Sort,a video based vehicle detection and tracking software system for emergency vehicle is developed,which can count the number of the vehicles and extract the trajectories of vehicles.The detection and tracking effect of the system is verified by road video test.