Research For Vehicle Detection And Tracking Algorithm Based On Single Camera

Vehicle detection and tracking has become a hot issue in the field of Intelligent Video Surveillance,which has been widely used in intelligent transportation,Traffic Control,Vehicle Tracking,etc.However,there are many problems remain to be solved in detection and tracking of moving vehicle for the complexity of scene.Therefore,our study focuses on the vehicle detection and tracking methods in complex scenes.We propose several new methods and achieve satisfying results in practical application.Our main work is as follows:a)With the further research of Gaussian Mixture Model(GMM),we analyzed the merits and demerits of object detection algorithm.Considering the actual situation of the scene,we proposed to adjust the learning rate of GMM model by changing the exchange frequency of foreground and background,which could meet the changes in different regions.b)We collected the realistic vehicle data set and trained the YOLO v2 model,which considered as the vehicle detection submodule in the tracking algorithm.The detection result rectify the bounding-box during tracking process when the attitude and scale of the vehicle changes.Therefore,the tracking box is more relevant to the real target for reducing the error rate of vehicle tracking.c)We compared the advantages and disadvantages of state-of-the-art tracking algorithms.We proposed a new strategy that the model in current frame exploited the vehicle position tracked in previous frame by the negative feedback method.To modify the vehicle position of current frame,we introduced the precise matching rate of feature points and the area overlap as the impact factors for the current area.d)We put the matched feature points into the container and compare it with the detected feature points of vehicle in order when the vehicle was occluded or the position offset is too large.The target vehicle with the highest matching rate will continue to track.Our algorithm can track the target vehicle again in the motion scene followed by the drone although the target vehicle drifts.e)We propose an integrated experimental system for the surveillance video in practical scene with satisfying result in vehicle detection and tracking.