Research On UAV Video Multi-object Tracking Method Based On Image Correction

The multi-object tracking task is to identify multiple types and numbers of objects in consecutive video frames and to assign the same objects the same number.The objects tracked in a multi-object tracking task are usually pedestrians and various vehicles.With the popularity of UAVs in commercial use,there is an increasing interest in analyzing the content of videos captured by UAVs.The multi-object tracking task of UAV videos has become a hot problem in computer vision for the need of person identification and tracking,and thus contributing to building smart cities.Traditional multi-object tracking methods are difficult to obtain accurate tracking results when facing problems such as motion blur caused by high-speed camera movement,inconsistent target scales,and coupling of target and camera motion.In this paper,we propose a multi-object tracking method based on image correction to address the above problems,and the research content is as follows.(1)To address the problem of low target detection accuracy in videos captured by UAVs,a multi-target tracking framework combining YOLOv5 and Deep Sort with first detection and then tracking is proposed,which can identify objects more accurately while reducing the false positive rate and making the tracking effect more stable.YOLOv5 algorithm fuses feature pyramid network to achieve multi-scale detection,and adopts data enhancement,adaptive bounding box calculation,adaptive graphical deflation and other methods to adapt to the problems of small objects and large changes in object scale in UAV filming;Deep Sort multi-object tracking algorithm’s Deep Sort multi-object tracking algorithm has become a widely used multi-object tracking paradigm because its multi-level matching mechanism can maximize the association of the same objects in adjacent frames.Combining the two can give full play to their own advantages and provide a basic framework for multi-object tracking of UAV videos.(2)For the problem of motion blur caused by high-speed camera movement,a hybrid deblurring strategy is proposed and used as a pre-processing module for object detection.Setting the deblurring encoder and generating a new sequence according to the blurring degree of each frame can improve the clarity of each frame sequence and enhance the multi-object tracking effect.Video blurring causes distortion of the object in some frames,which cannot be recognized and leads to broken tracking trajectories.In this paper,we propose deblurring the video to improve the success rate of object detection.The principle of deblurring is to infer the optical flow information based on the features of the preceding and following frames,and to repair the parts that may cause blurring.To improve this situation,a hybrid deblurring strategy is proposed,in which a deblurring encoder is set to filter out frames that need to be deblurred,and these frames are combined with clear frames to form a new sequence after the deblurring process.Experiments prove that the hybrid deblurring strategy can better improve the multi-object tracking effect.(3)For the motion coupling problem of camera and object,a global motion compensation method is proposed,which can align the feature points of two adjacent frames,decouple the camera motion from the object motion,and improve the accuracy of multi-object tracking.Different from the traditional multi-object tracking dataset,the video captured by the UAV not only has the high-speed motion of the object,but the UAV carries a camera that also performs irregular motion combinations such as backand-forth movement,rotation,lift-off or landing at high speed.The motion coupling between the camera and the object leads to errors when linearly predicting the position of the object in the next frame in traditional multi-object tracking,resulting in object re-identification failure.To this end,this paper proposes to use feature matching to obtain the single-strain conversion matrix of adjacent frames,through which the feature points of two adjacent frames can be aligned to cancel the global motion of the camera,thus achieving the effect of decoupling the motion and making the linear motion prediction more accurate.Experiments prove that the global motion compensation method can improve the accuracy of multi-object tracking.In this paper,the improvement in multi-object tracking accuracy of the above methods is tested in the Vis Drone dataset,respectively.Based on the two-step multiobject tracking methods using YOLOv5 and Deep Sort,this paper proposes a multiobject tracking framework incorporating two image correction methods of hybrid deblurring and global motion compensation,and compares the performance differences with other algorithms on the Vis Drone dataset,and analyzes the experimental results.The experiments prove that the proposed method achieves better tracking results in this paper.