Research On The Real Time Vehicle Tracking Based On Unmanned Aerial Vehicle

Visual-based vehicle tracking is one of the hot issues in modern transportation system.Its main task is to detect and track the vehicle in the surveillance video and obtain the traffic information of the road,which can provide reference information for further rational allocation of traffic resources.Compared with the traditional fixed-angle video capture,the use of rotor-unmanned aerial vehicles(UAVs)to monitor and track vehicle can break through the restrictions of perspective and achieve a wider range of traffic information collection.Therefore,it is of great value to study and design an efficient and accurate UAV vehicle tracking algorithm.In recent years,with high tracking accuracy and fast processing speed,the popular visual tracking algorithm Kernel Correlation Filter(KCF)is a good choice for real-time vehicle tracking for UAVs.However,due to the occlusion and relative motion between the UAV and the target,the algorithm can not fully deal with practical problems such as fast motion,motion blur and occlusion.In view of the above problem,we analyse the shortcoming of the KCF and put forward corresponding methods to improve the tracking robustness.The main work of this paper is in the following areas:(1)There is a boundary effect in the process of circulant dense sampling in the KCF,which limits the search area of the tracker and may lead to the failure of the vehicle in fast motion.We propose a kernelized correlation filter tracking method based on the twice search for this limitation.In this paper,another kernelized correlation filter is added to model the target context under the framework of the KCF.The filter performs as a candidate model and provides a new detection point for the original tracker,which can expand the search of the tracker range.In addition,we also add a common scale estimation method.Experiments on public data set OTB 2013 show that the improved algorithm is 7.2% higher than the KCF in the whole average distance accuracy rate,and the accuracy rate under the attributes of deformation,fast motion,motion blur and background clutter is much higher.(2)A kernelized correlation filter vehicle tracking method based on Kalman filter is proposed to solve the problem that there is no occlusion judgment mechanism in the KCF.This method uses the output of the kernelized correlation filter to determine whether the target is occluded.Then it uses the Kalman filter in the case of occlusion to estimate the position of the target and searches the target over a wide range by the improved re-detection method.Experiments in the actual scene show that Kalman filter can estimate the position of the linear moving object accurately and improve the performance of the tracker when occlusion occurs.(3)Aiming at the complex occlusion problem encountered in the tracking process,a method of map-assisted tracking is proposed by means of sensor resources available on the UAV.The method pre-marks the occlusion information on the offline map and provides reference information for the tracker in the case of complex occlusion to improve the re-searching of the target.Experiments in actual scenes show that the use of off-line maps can effectively narrow the searching range of the tracker and increase the performance of the tracker in coping with complex occlusion.