Research On Multi-camera Pedestrian Tracking

Pedestrian detection and tracking is a very important topic in computer vision.However,due to the occlusion and the limited perspective of the camera,single-camera pedestrian tracking faces enormous challenges.Tracking with multiple cameras allows you to observe pedestrians from multiple perspectives,so it provide richer visual information to solve these problems.This article proposes a multi-camera pedestrian tracking system.The flow of the system is as follows:1.Acquisition of pedestrians using GOPRO camera.The video information is transmitted to the processing center.2 pedestrian detection is performed on the video captured by each camera in order to get the features of pedestrians.3.Pedestrians are tracked according to the results of pedestrian detection,and the information of the tracking objects is obtained.4.Matching and feature fusion of pedestrians from multi-camera and eventually get the pedestrian's trajectory on the common planePedestrian detection algorithm based on neural network can achieve real-time pedestrian detection result.Next,we apply the Markov decision process to model the state of the tracking object,which is divided into the initialization state,the tracking state,the lost state,and the termination state.The features of the tracking object and the detection object is extracted each frame.then we calculate the similarity of them and perform the maximum matching of the bipartite graph.After the matching between the tracking object and the detection target,the MDP status,the image and physical features of the tracking object will be updated.For the case that the tracking object does not match with any detection object,we will use GAN methods to generate the image of the current frame according to the first few frames of the tracking object,thereby updating the image feature and physical feature of it.The last step is the matching and fusion of pedestrians under multiple cameras.We project the tracking objects from the camera plane onto the common plane according to the homography,and then match the tracking objects from different cameras.The matching is performs mainly according to the location information and image features of the tracking object.After the tracking objects belonging to the same pedestrian are correctly matched,we can get the trajectory of the pedestrians on the common plane...