Study On Visual Tracking Algorithms Based On Spatio-temporal Correlation Filters

Visual tracking is one of the important technologies in computer vision.It has important applications in the fields of intelligent surveillance,military,human-computer interaction,intelligent transportation and so on.Target tracking task is to accurately track any target by self-adapting modeling the appearance model of the target after obtaining the initial position annotation information of the target.Because the mode of this kind of online tracking task is not fixed,it faces many challenges,such as illumination change,target appearance change,system change,low resolution,occlusion and target moving out of view,etc.Therefore,designing a tracking algorithm with high accuracy,strong robustness and high computational efficiency has always been an important research topic.The tracking algorithm based on correlation filter has high accuracy and fast calculation speed and has dominated the tracking algorithm in recent years.However,the whole training process of the filter does not model the background information,but only uses the periodic extension of the samples,so it does not take into account the spatial information of the whole image.In addition,the tracking process is a continuous image sequence in the spatial domain,and the traditional correlation filtering pays little attention to the temporal consistency information between frames,which leads to the tracking failure.Based on the algorithm of correlation filter,aiming at the problems of occlusion,target moving out of view,background mixing,aspect ratio change and target fast motion,our work propose the improvement of traditional correlation filter tracking method in feature learning,motion model design and inference strategy focusing on temporal and spatial information,in order to obtain high accuracy and real-time algorithm.The main contributions of this paper are as follows:Aiming at the problem of tracking failure caused by background cluster,occlusion and moving out of view in long-term tracking,a temporal-spatial correlation filtering tracking algorithm based on foreground-aware is proposed in this thesis based on spatial regularized correlation filtering(SRDCF)tracking algorithm.In the tracking task,the correlation filter is designed for the actual information of background and foreground to make the correlation response more robust.A time consistency model is proposed,which considers the context information in tracking tasks and prevents the sudden change of filter template.In addition,in the process of long-term tracking,a re-detection method is proposed.By judging the tracking drift,the detector is activated to retrieve the target,so as to achieve the effect of long-term tracking.The algorithm is validated by OTB data set,and the tracking accuracy and success rate are higher than SRDCF.It can effectively deal with occlusion,target moving out of view and other scenarios,and can realize online tracking under the expression of traditional features.Aiming at the problem that correlation filter algorithm has poor robustness to fast motion and change of aspect ratio,a spatio-temporal correlation filter tracking algorithm based on image change perception learning is designed in this thesis.In this work,the correlation filter calculation is transformed into a layer of the network to learn the feature extraction part of the convolution neural network through back propagation,which enhances the discriminant ability of the model.Secondly,by introducing the spatial transformation network,the position of the target is roughly estimated by regression method,and the motion parameters of the target are estimated at the same time,so as to express the spatial-temporal correlation of the image.The image change perception learning module is cascaded with the correlaton filter network module,and the end-to-end off-line training is carried out to enhance the generalization ability of the network.Through off-line training,the tracking algorithm only updates the filter parameters while tracking,and realizes on-line tracking.The output of the network is fed back to the input to realize the process of circular tracking.The validity of the algorithm is verified on OTB dataset,and the results show that the algorithm has good tracking effect for many different scenarios.