Research On Pedestrian Detection In Congested Scenes Based On Improved YOLOv5

Pedestrian detection plays an important role in scenarios such as automatic driving and security monitoring,and is an important research direction in the field of computer vision.In pedestrian detection tasks,object detection algorithms based on deep learning are widely used,but the performance of pedestrian detection in crowded scenes still needs to be further improved.In addition,the target detection algorithm has been continuously improved,the performance has been continuously improved,and the volume of the model,the amount of parameters and the amount of calculation have also been continuously increased,which puts forward higher requirements for the performance of the hardware.Therefore,this article focuses on the application of YOLOv5（You Only Look Once version5）algorithm in pedestrian detection in congested scenes,which has low accuracy and large model volume and parameter quantity.The main work is as follows:（1）A parameter optimized YOLOv5 pedestrian detection algorithm for congested scenes has been proposed,which improves the accuracy of YOLOv5 algorithm in pedestrian detection in congested scenes.The K-means++algorithm is used to cluster the marked frames of pedestrians in the dataset,so that the prior frames are more suitable for pedestrian targets of different sizes and proportions,and the accuracy of model prediction is improved;the bounding box regression loss function is improved to make the regression of bounding boxes more accurate.Precise,so that the model can locate and detect the target more accurately;use the merge non-maximum suppression algorithm（Merge Non-Maximum Suppression,Merge-NMS）to filter the prediction frame,so that the filtered prediction frame is more accurate and more accurate.A lot of target information.The Recall,mAP,and mAP_50:95 of the improved algorithm reached 0.787,0.863,and 0.549,respectively,which were 4.8%,1.3%,and 3.6%higher than the original YOLOv5 algorithm.（2）A YOLOv5 pedestrian detection algorithm for congested scenes based on GhostNet convolution optimization has been proposed,which reduces the model volume and parameter quantity of YOLOv5 algorithm applied to pedestrian detection in congested scenes.Use GhostNet convolution as the three convolutional layers at the output end of the network to reduce the volume and parameters of the model;introduce EIoU Loss（Efficient IoU Loss）to make the error calculation between the predicted frame and the real frame more reasonable and improve the detection performance of the model.The Recall,mAP,and mAP_50:95 of the improved algorithm reached 0.766,0.863,and 0.535,respectively.Compared with the original YOLOv5 algorithm,they increased by 1.6%,1.3%,and 0.9%,respectively,and the detection speed increased from 39fps to 41fps.The model volume It has decreased by 5.2%,and the number of parameters has decreased by 5.4%.In this paper,aiming at improving the accuracy of pedestrian detection in crowded scenes and reducing the model volume and parameter quantity,based on the YOLOv5 algorithm,a series of improved methods are used to conduct experiments on the public Widerperson dataset.The experimental results prove that the improved algorithm is equivalent to It is more suitable for pedestrian detection tasks in crowded scenes than the original YOLOv5 algorithm.