Research On Pedestrians Detection Algorithm In A Crowd

Pedestrian detection is a vital branch in the field of computer vision,which mainly refers to finding pedestrian targets and determining their positions from massive images or video frames.It is a challenging application scenario in pedestrian detection in-crowd.The challenge is mainly reflected in the high target density and the self-occlusion between pedestrians,which leads to the inaccuracy of the boundary box position of the loss function regression.The hard sieve of the post-processing of the detection framework has caused missed and false detection of the target,which resulting in low accuracy of pedestrian detection in dense scenes.With the rapid development of artificial intelligence technology and computer hardware in recent years,the target detection algorithm based on deep learning has achieved excellent performance in detection accuracy and has a target detection algorithm based on deep learning is widely used in many scenarios such as intelligent unmanned driving,regional security monitoring,and intelligent transportation,etc.To meet the application of the detection framework in engineering and solve the above problems.We are based on the deep learning-based end-to-end object detection algorithm YOLOv3.The main research work is as follows:1.This paper uses the YOLOv3 algorithm of multi-target detection as the detection framework.Given the difference between the shallow and deep features of the convolutional neural network,the corresponding receptive fields of different feature layers are also different.The deep feature map contains a lot of semantic information and the shallow feature map contains more geometric information.We propose a p-YOLOv3 network to prevent the errors caused by direct stitching between feature maps of different depths.In the network,the "131" convolution structure is used to extract and recombine the shallow features,the convolution structure features and deep features are spliced,and then the original detection network is cut to extract the effective feature information of the target pedestrian.The experimental results show that the p-YOLOv3 network architecture has more advantages than the original network in the pedestrian detection data set.2.To obtain the accurate position information of the target,a reasonable loss function is set up in the model training to make the prediction frame close to the real target frame.However,in dense scenes,the density of real frames is large,and there is an impactbetween frames.To make the prediction box close to the real matching box and away from other surrounding real target boxes.This article added the repulsive force loss to the loss function of YOLOv3.In model is training,the prediction box is punished when it is close to other unmatched real boxes,and a selection strategy is set for other unmatched real boxes.Experiments show that the loss function after adding repulsive force is helpful for bounding box regression,and thus improves the detection performance of the detection network in dense crowds.3.In the post-processing part of the target detection network,for the prediction frame with higher confidence,it is only determined whether to keep the prediction frame by calculating the overlap with the prediction frame with the maximum confidence.This method of filtering the prediction frame pedestrian detection in the scene is very unfavorable.In order to keep the effective prediction frame,we have improved the postprocessing mechanism,used Generalized Intersection over Union(GIOU)to judge the similarity between the two prediction frames,and the secondary prediction frames with high similarity and to be filtered out are screened.The second screening is mainly to calculate the distance between the prediction frames.By comparing the distance and the distance threshold,the prediction frame is left to be determined.Experimental data show that the improved post-processing method can effectively retain the prediction frame of the real target and improve the average accuracy of the detection model.