基于改进的YOLOv5算法道路目标检测分类技术研究

Research on road target detection and classification technology based on improved YOLOv5 algorithm

随着互联网技术的飞速发展,智慧交通系统(Intelligent Traffic Systems,ITS)技术日渐成熟。其中,智慧道路交通管理服务领域需要对交通动态信息进行实时检测,在获得相应路口车流量信息的同时,还需要对经过该路口的目标进行分类、统计、追踪。因此,如何对经过路口的目标进行检测分类逐渐变成当前视觉领域研究的热点问题。该文借助雷视一体机收集了大量道路目标数据集,利用处理后的数据集在pytorch框架上进行训练。通过标签平滑处理、数据增强、改进损失函数等方法来提升检测效果。实验对比之后,该模型处理结果比优化前的模型对远处小目标、遮挡车辆、镜头前大车辆等目标的识别效果更好,mAP达到了77.57%,相比原先的网络提高了2.76%。

With the rapid development of Internet technology,Intelligent Traffic Systems(ITS) technology is becoming more and more mature. Among them,the intelligent road traffic management service field needs to detect the traffic dynamic information in real time. While obtaining the traffic flow information of the corresponding intersection,it also needs to classify,count and track the targets passing through the intersection. Therefore,how to detect and classify the targets passing through the intersection has gradually become a hot issue in the field of vision. In this paper,a large number of road target data sets are collected with the help of mine vision all-in-one machine,and the processed data sets are trained on the pytorch framework. The detection effect is improved by label smoothing,data enhancement and loss function. After experimental comparison,the processing result of the model is better than that of the optimized model for the recognition of distant small targets,occluded vehicles,large vehicles in front of the lens and other targets,mAP reaches 77.57%,which is 2.76% higher than the original network.