Research And Application Of Target Pedestrian Tracking In Multi-camera System Environment Based On Deep Learning

With the gradual increase in the coverage area of video surveillance,the frequency of pedestrian appearance has gradually increased.At present,the detection,tracking,and re-identification of pedestrians through the use of video surveillance data and data analysis methods have become the main technical means of intelligent transportation and security systems.However,in real scenes,the range of pedestrian activities exists in multiple cameras,and there are also problems such as viewing angle,illumination,posture,background clutter and occlusion,which makes it impossible to accurately and quickly detect,track and re-identify pedestrians.Therefore,multi-camera pedestrian tracking has become a current research hotspot.First of all,in terms of pedestrian detection,long-distance and small-sized pedestrians are difficult to detect,and they are easily interfered by complex backgrounds.In addition to the problems of low resolution and less effective information,an improved method combining faster region-convolutional neural network(Faster R-CNN)to detect long-distance and small-sized pedestrians in a complex background is proposed.The mixed Gaussian model is used to solve the problem of complex background interference.It can remove the background information of the video in time and extract the foreground of the video image.In order to further solve the problem of low resolution and less effective information,the bilinear quadratic interpolation method is used to enhance the resolution of the image,and multi-scale feature fusion is used to make up for the lack of effective information.Experimental results show that the detection accuracy of pedestrians in different scenarios has been improved,and the detection accuracy of long-distance and small-sized pedestrians has been improved more significantly.Secondly,in terms of pedestrian tracking,pedestrians always have a similar appearance,and there are occlusions,intersecting trajectories in a crowded scene.There may also be problems such as easy to lose data and camera movement.A multi-pedestrian online adaptive tracking algorithm based on you only look once(YOLO)is proposed.YOLO has fast tracking speed and can track pedestrians in crowded scenes.In view of the problem that the continuous storage of video streams will occupy a large amount of memory space,the dynamic sample set method is adopted to ensure that there is enough storage space to store the samples,and to strengthen the real-time tracking.Considering that the direction of pedestrian movement is nonlinear in most cases,a Markov chain is adopted to fit a small trajectory with a relatively high degree of approximation to the complete trajectory of the pedestrian.Finally,the method proposed in this paper is verified on the PETS2009 data set and TUD data set,which proves the rationality of the algorithm proposed in this paper.Finally,in terms of pedestrian re-identification,as traditional pedestrian re-identification algorithms are susceptible to different perspectives,illumination,posture,background clutter and occlusion,a pedestrian re-identification method based on joint segmentation is proposed.In the key point segmentation,the joint segmentation method is adopted to divide the pedestrian as a whole into global features and seven local features.In metric learning,the weighted triplet(WT)loss function is used to make positive examples closer to the anchor point,and negative examples farther away from the anchor point.Training and testing are performed on the Market-1501 data set.The experimental results show that the Rank-1 and m AP of the proposed method are increased to 88.6% and 72.4% respectively.It has achieved better detection results in a multi-camera system environment.Finally,based on the multi-view environment,the fusion of multi-view and single-view is compared and analyzed.The results show that the fusion of camera views helps to improve the detection results.