Research On Pedestrian Detection And Video Enrichment Based On Deep Learning

In recent years,with the rapid development of deep learning,pedestrian detection and video enrichment have become the focus of research in the field of video surveillance analysis.Pedestrian detection is a basic study in computer vision.Its main task is to analyze whether the target pedestrian is included in the video image and accurately locate the pedestrian.Video enrichment technology is to concentrate the main content in the original video into a concentrated video that is convenient for people to browse.It plays an irreplaceable role in maintaining social security in the field of social public safety.The main research contents of this paper are as follows:At present,the real-time detection of pedestrian algorithm is not ideal in detecting the occlusion of the crowd.To solve this problem,the repulsion loss function is applied to Faster R-CNN for pedestrian detection.The repulsion loss function consists of two factors: one is the mutual attraction factor between pedestrians,and the other is the repulsion factor with other surrounding pedestrians.The repulsion factor prevents pedestrian candidates from moving to surrounding pedestrians,making pedestrian detection in the crowd more robust.The experimental results show that the pedestrian detection algorithm using the repulsion loss training has a significant effect in crowd pedestrian detection,and a higher detection accuracy is obtained.Aiming at the problems of low accuracy and high rate of missed detection in pedestrian detection in realistic complex environments,image HOG features combined with deep learning features are applied to pedestrian detection.Firstly,the pedestrian edge description operator HOG feature and pedestrian depth semantic feature are obtained by statistically analyzing the gradient information of the pixel points in the image and using the ZF-Net feature generation network.Then,the candidate region generation network is used to process the two features.Pedestrian candidate regions of various scales and aspect ratios are output.Finally,the two features and pedestrian candidate regions are processed using a Fast R-CNN network.Theexperimental results on the INRIA and Caltech datasets show that compared with the current mainstream algorithms,the proposed pedestrian detection algorithm can successfully detect pedestrians in complex background situations.Aiming at the problem of mutual occlusion and background complexity between moving objects during video concentration,it is difficult to accurately extract moving objects to cause the concentration ratio to decrease.A convolutional dual-stream fusion neural network video concentrating method based on interaction mechanism is proposed.Firstly,the region of interest is selected for the input video frame.Then,the convolutional dual-stream fusion neural network is used to extract the moving object features and background features and perform feature fusion to reduce the influence of mutual occlusion between the moving objects.Finally,the fused features are correlated by the interaction mechanism to effectively improve the correlation between the moving objects and the moving objects and the background,and then the scene is clustered according to the similarity matrix to obtain the key frames.The experimental results show that the video concentration through the network structure,the concentration ratio and the recall rate are improved.