Research On Pedestrian Detection Algorithm Based On Improved YOLOv5

With the continuous development of the national economy,cars enter every home gradually,and traffic safety issues have also become the focus of research.The emergence of the driver assistance system has greatly alleviated the harm caused by traffic safety problems.It can not only help the driver when driving,avoid the traffic safety problems caused by the driver’s carelessness as much as possible,but also assist the driver in all directions when braking,backing up,and parking,greatly easing the pressure on the driver when driving and improving the safety.Compared to drivers,pedestrians are often more seriously injured in traffic accidents.Therefore,the detection and identification of pedestrians have become a very important technology in the driver’s auxiliary system.The inspection and identification of pedestrians allow the vehicle to perceive the existence of pedestrians as soon as possible and avoid or relieve the risk of conflict between people and vehicles.These are effective measures to protect pedestrians and reduce casualties.This article is aimed at the problems of missed inspections and wrong inspections due to the obstruction of pedestrians during the inspection of pedestrians.A pedestrian test algorithm based on YOLOv5 improved.The specific work of this article includes the following aspects.(1)For pedestrian testing due to pedestrian density.The background image is very similar to pedestrians visually.At this time,the objects in the background are often predicted as pedestrians.Replace some of the C3 modules in the Backbone and Neck structure of YOLOv5 with Swin Transformer Tiny’s self-attention mechanism module.This kind of self-attention mechanism is based on a sliding window.The characteristics of the sliding window can be used to communicate between the windows of each part to achieve the effect of global modeling.Introducing context information can effectively improve the error classification of difficult samples such as the background.(2)For the target objects of overlap blocking,their convergence speed is often slow,and the problem of missed inspection may occur.The loss function in YOLOv5 is replaced from the original CIo U Loss to SIo U Loss.The SIo U Loss used considers the angle between the real box and the prediction box.It helps reduce the missed inspection problems existing during pedestrian testing and improves accuracy.(3)To overcome the problems of approaching and even overlapping,it may cause the problem of filtering the prediction box of another pedestrian.Use a confidentbased non-maximum suppression CP-Cluster.This method uses the relationship between the candidate box to spread messages between the candidate boxes to adjust its confidence.And the confidence spread iteration many times so that each candidate box is fully parallel.This method can effectively screen the correct prediction box,try to avoid leakage,and can get a performance improvement.(4)Verify the effectiveness of this article by designing multiple groups of ablation experiments.Compare the experiment with the current mainstream pedestrian test algorithm and analyze the results of the experiment.Verify the effectiveness of the algorithm used in this article.Considering the actual application level of the driver’s auxiliary system.The improved YOLOV5 algorithm will also be improved.Use the NCNN deep learning framework to deploy it to the mobile Andriod.Provide help for the real application in the development of smart transportation.