Research Of UAV Multi-object Tracking

With the rapid development of microelectronics technology,the performance and endurance of unmanned aerial vehicles（UAVs）have been rapidly improved.In addition,due to its high flexibility and high adaptability,UAVs have become an irreplaceable and important means in many application scenarios.In particular,the employ of visual multi-object tracking technology on the UAV platform can greatly improve the intelligence and automation of UAVs.This enables the UAV to automatically analyze the scene information quickly and efficiently during the execution of tasks,and help people make decisions in the ever-changing work scene.However,with the high complexity and variety of UAVs working scenarios,applying multi-object tracking algorithms to UAV missions faces some challenges:（1）Challenges brought by the flight characteristics of UAVs.The high maneuverability of the UAVs make the camera it carries in rapid motion.At the same time,during the high-altitude flight of the UAVs,inevitable mechanical vibration will occur,resulting in motion blur in the images captured by the drone.（2）Challenges in the UAV object tracking process.The high-altitude shooting angle of the UAV results in dense targets in the field of view,and the visual changes and target occlusion brought by the flight process will cause the model learned by the target tracking method drift easily.（3）Challenges brought by the limitations of the UAV platform itself.Most UAVs are only equipped with one CPU,which computing resources are lacking.Besides,the cameras are prone to image blur and noise under the influence of the external environment,resulting in image degradation.Focusing on the target tracking technology based on UAV,this paper proposes a series of new models and methods for the problem that the existing tracking algorithms performance degradation in UAV applications,so as to improve the performance of multi-target tracking in UAV applications.1)A UAV parallel dual network tracking framework to cope with occlusion.In this section,a parallel dual network tracking framework that can identify tracking drift and relocate the tracking target is constructed for the frequent occlusion and perspective changes in UAV application scenarios.By exploring the advantages and disadvantages of the Siamese tracking method and the correlation filter tracking method in the target tracking process.We construct a model drift discrimination mechanism,when the tracking drift is caused by occlusion and viewing angle changes.This model can identified and processed abnormality in time,so that the tracking method can be re-tracking.This tracking framework ensure that the tracking failure due to occlusion and visual changes can be corrected and resumed in time in the UAV application.On the UAV123 dataset,our tracking framework relocates the object after occlusion and thus achieves tracking precision and success rate of 0.765 and 0.635.2)Background multi-cue suppression and aberrances response supperssion mechanism for UAV.Aiming at the problem that the target in the UAV image is too small and the effective information that can be used is lacking,we have studied and explored a correlation filter with the ability to perceive background information.By extracting information from four dimensions,the full use of image information by the tracking method is increased,and the discrimination ability of the tracking method is improved.Motion blur and the use of too much background information can add noise to the algorithm.We adopt an aberrances suppression mechanism to ensure that the learned target model has a smooth consistency,which ultimately improves the robustness of the tracking method on the UAV platform.On the UAV123 dataset,our method improves the precision and success rate by10.2% and 10.7%,respectively,compared to the baseline tracker without anomaly suppression mechanism.3)UAV multi-object tracking algorithm with multiple judgment mechanisms.In order to achieve a high-precision and robust UAV multi-target tracking algorithm,this paper proposes a multi-object tracking framework with multiple judgment mechanisms,which consisting of a backbone network and three auxiliary modules.It adopts a backbone network that can adapt to targets of different sizes,fully extracts information from the targets identified by the detection module,and distinguishes difficult and easy targets by judging the effective information of the targets.Then,the data association module is used to quickly assign the trajectory to the easy target,while the difficult target adopts the tracking module based on correlation filtering and Siamese network to assign the trajectory,and finally realizes the multi-target tracking method of the UAV scene.Finally,on the Vis Drone2019 dataset,MOTA and MOTP scores of 33.1 and 78.4 were obtained while maintaining real-time performance.