Research On Object Tracking Method Via Spatial Information Learning Under Complex Environment

Object tracking is such an algorithm to locate the target continuously in the video frames when given a continuous image sequence and the position of the interested target in the initial frame.Object tracking plays an important role in the task of machine vision because of the natural requirements of real-time positioning the object,which includes automatic driving,intelligent security,UAV application,satellite remote sensing and many other important application fields in the national economy.In recent decades,the research of object tracking has attracted the extensive interest of researchers,although the existing tracking methods have already overcome some difficulties in this field and made great breakthroughs,the spatial context around the target object has brought many disturbances in the complex environment,which has caused many problems to be solved,these problems currently restricting the performance of the object tracking algorithms to be further improved.Current object tracking methods only use a single filter when determining the target's position,and after being disturbed by the surrounding information in a complex environment,the tracker lacks effective spatial reference information,which may easily cause the location drift under the blind decision of the single tracker.At the same time,during the moving process,the target object is inevitably disturbed by serious occlusion or dramatic changes in light,which will significantly change the appearance model of the object.The existing tracking algorithms urgently need a simple and effective way of the method to deal with such a scenario,so that the tracker can detect such a situation and deal with it after the appearance of the object has changed dramatically to enhance the robustness of the tracking algorithm.Inspired by the image classification task,the current tracking algorithms mainly learn both positive and negative samples to obtain the ability to distinguish between target object and background,which has weak adaptability in complex scenes and may easily cause the tracker to drift or even lose the target.On the other hand,to make full use of the ability of reinforcement learning to describe the motion information of the object in the video,some researchers began to study the target tracking algorithm based on the deep reinforcement learning.However,in a complex environment,the scale of the target object may change at any arbitrary on the two axes of the video screen.Current deep reinforcement learning-based tracking algorithms can not estimate the scale of the object correctly in a complex environment precisely.A better approach is needed to complement the shortage of current tracking methods in this regard.In this paper,aiming at the above several problems in the field of target tracking,we study them from the perspective of spatial information learning,proposing corresponding solutions and obtain some valuable research results.Firstly,a correlation filter tracking method with multi-scale spatial co-location is proposed.In the proposed algorithm,the tracker contains a total of three sub filters with different sampling ranges,and this paper also proposes a collaborative discrimination method for positioning according to the spatial response of the target image sample of each filter,and jointly locates the target online.Based on the proposed method,this paper further explores the potential power of the correlation filter tracking algorithm in a complex environment and proposes a robust correlation filter tracking method with multi-scale spatial view.The proposed method can detect whether an object is seriously disturbed,including severe occlusion and dramatic changes in illumination.In addition,the target state memory mechanism is also set up in the algorithm to restore the appearance of the object in time when it is detected to be seriously disturbed.A large number of experiments on the OTB tracking benchmark show that the proposed algorithm has a well performance.In order to solve the shortcomings of current target tracking algorithms in sampling methods,this paper proposes a metric correlation Siamese network and multi-class negative sampling method for visual tracking.In the aspect of sampling,the proposed algorithm divides negative samples into three categories to better adapt to target tracking tasks.In addition,this paper uses the idea of metric learning combined with the ridge regression method in the Fourier domain,a metric correlation module is proposed,which enables the filter better learn the discrimination information of the target and background.A large number of experiments on OTB and VOT target tracking benchmarks show that the performance of the proposed algorithm is ahead of the current mainstream tracking algorithm.In order to solve the problem that the current deep reinforcement learning-based tracking algorithms are hard to correctly estimate the pixel aspect ratio of the target in a complex environment,this paper proposes to use deep reinforcement learning to determine the center position of the target object and introduces Io U estimation network to achieve bounding-box prediction.In addition,to suppress the background noise in a complex environment,this paper introduces the deep learning-based segmentation in the proposed algorithm and combines the prior space fusion to form the final spatial attention for better adapting to the deformation produced by the object during its motion.This paper has conducted extensive experiments on the OTB,UAV123,and VOT benchmarks,the results show that the proposed algorithm achieves to the state-of-the-art tracking performance.