Research On Vehicle Tracking Under UAV Scene Based On Kernel Correlation Filtering

With the increasing number of vehicles,road traffic problems are closely affecting daily life.Traffic regulators need to monitor road condition information in real time through video data.The rapid development of drone technology has made it possible to obtain data from an aerial perspective.UAVs can provide a broader and more flexible view,contain more information,and are also capable of long-distance surveillance.Tracking vehicle targets by images and videos under UAV scenes is also expected to be a new way of traffic monitoring.The tracking scene under UAV view is not stable enough,the target is prone to substantial deformation and may be obscured by the environment.The hardware platform of the UAV also determines that the algorithm cannot perform overly complex operations to achieve real-time.Therefore,the ability to achieve fast and accurate vehicle tracking in complex environments with lightweight embedded devices will have important research value and practical significance.In this paper,after extensive research and study of various target tracking algorithms that have been proposed so far,the kernel correlation filtering algorithm is chosen as the main object of study.It is because it can achieve a faster operation speed on the embedded platform,at the same time the accuracy can be maintained at a high level.Based on this algorithm,the corresponding optimization strategy is proposed by combining the difficulties of vehicle tracking in UAV scenarios with the scenarios of practical applications.The main research of this paper has the following three parts.(1)In the UAV scene,the traditional nuclear-related filter tracking algorithm is prone to track frame offset when facing fast-changing targets.In order to solve this problem,this chapter proposes a scale-adaptive tracking algorithm that incorporates color features.A spatial tracker with different scales is designed to synchronize the location and size estimation of the target vehicle,which can quickly and accurately determine the relevant information of the target vehicle after deformation,and realize the self-adaptation of the scale.In addition,in order to increase the robustness of the algorithm,color features that are not sensitive to deformation are also added,and the statistical color feature method is used,which is not limited by template features,effectively coping with the problem of scale mutation of the target vehicle in this scene.The test results on the data set show that the improved algorithm can effectively cope with the target scale change,the tracking accuracy is improved compared with the traditional algorithm,and the tracking speed has also reached the real-time standard.(2)Aiming at the problem of template pollution when encountering similar target interference and severe occlusion in the long-term vehicle tracking in the UAV scene,a longterm vehicle tracking algorithm based on high-confidence model update is proposed.The template is no longer updated every frame,but by calculating the similarity between the current template and the tracked vehicle,it is used to determine whether the current template can reflect the characteristics of the target,so as to decide whether to update or not.Among them,the classifier uses a structured support vector machine combined with correlation filtering,which improves the speed of the algorithm.When encountering similar vehicle interference,multipeak detection is used to avoid possible model drift.The maximum value of the response graph and the average peak-to-peak correlation energy index are used to determine whether to update the current learning template.The test results on the data set show that the improved algorithm can effectively deal with the problem of long-term tracking of the target being occluded,the tracking accuracy is improved compared with the traditional algorithm,and the tracking speed also reaches the real-time standard.(3)Based on the tracking algorithm proposed in this paper,an intelligent vehicle tracking system is designed,which can accurately track the target vehicle in the video,and at the same time distinguish the occlusion and loss of the target vehicle.This system is developed on embedded equipment for the UAV scene.The test results show that the intelligent vehicle tracking system designed in this chapter can achieve real-time performance on the embedded platform and maintain a certain accuracy rate,which has certain application value.