Research On Object Tracking Merged With Handcraft And Deep Features

In recent years,among all the visual object tracking methods,particle filters based on handcraft features of image surface texture and color names,and deep learning trackers based on large samples' data training are the two critical research areas.Each of them develops many marvelous tracking algorithms.Handcraft features are sensitive to image textures,and have high precision of location and high computing speed,but lack of the adaption of excessive deformation and occlusion,which means lack of robustness.Deep features obtained by deep learning have high-dimension semantic information and are invariant to deformation and rotation,which lead to more accurate discrimination of the shape of the target,but lack of the adaption of high precision of movement.Besides,most of those popular object tracking algorithms are tend to get the position and scale estimation of the target faster and more accurate under mild background interference,but use simple strategy to deal with the problems of target re-search for being hid for a long time or coming out after disappear,so they are not qualified for long-term tracking.Based on those situations,this paper deeply analyzes the basic knowledge and those popular algorithms of visual object tracking,and also studies the principals of KCF(Kernel Correlation Filter)and DSST(Discriminative Scale Space Tracking),the filter with Spatial Reliability and PCA-HOG(Principal Component Analysis of Histogram of Oriented Gradient)Features and their implementation for short-term tracking,the methods of response strategy for merging handcraft and deep features,the target re-search strategy by the best target templates' matching and their implementation for long-term tracking.The details are as following:1)The basic knowledge and analysis of visual object tracking.This paper lists the difficulties of visual object tracking,including illuminationchanges,camera motion,target deformation,occlusion,background interference and scale changes.In addition,the current research status of visual tracking technology at home and abroad have been briefly described.2)Introducing the basic principle of the classic visual object tracking.This paper discusses the details of two classic object tracking algorithms,KCF and DSST.In order to solve the problems of boundary effect and the sample training distortion caused by the circulant matrices operator in KCF,and also the problem of the large calculation for the HOG features,the advance edition of CSR-DCF(Discriminative Correlation Filter Tracker with Channel and Spatial Reliability)is proposed.In this paper,the PCA-HOG features dimensionally decreased by PCA method and the color names features are introduced to fully describe the texture information of image.Besides,the spatial reliability mask is used to improve the probability of target area recognition and weaken the dependence on the shape of the target.The scale correlation filter also utilizes the PCA-HOG features to implement a scale description while the target's size is changing.In this paper,this algorithm is used as a short-term tracking strategy.3)The visual object tracking algorithm with feature response fusion of handcraft and deep features.In the process of handcraft features,the factorized convolution operator simplifies the target model and accelerate the calculations.The GMM(Gaussian Mixture Model)is used as the strategy of management the positive samples in order to effectively distinguish the target and the background.The sample data enhancement and incremental learning algorithms are used in the deep features processing,which train the sample model to enhance the generalization ability.The deep network structure uses a combination of Siamese Network and the RPN(region proposal sub-network)to improve the ability of parallel computing.For the strategy of the merging algorithm,consider the target position and scale description of the prior frame as the benchmark,so that the two types of tracking algorithms operate independently,and the feature response matrix obtained in the target search region is linearly planned to get the optimal weight assignment,andfinally get the optimal target response.For the target research strategy,consider enabling the strategy when target response value is below the specified threshold,and correcting the target misjudgment according to the best sample group,so as to achieve the target capture after occlusion or disappearance.In this paper,this algorithm is used as a long-term tracking strategy.4)The comparison of short-term tracking and long-term tracking algorithms.The target tracking algorithm based on spatial confidence mask and PCA-HOG feature is used as short-term tracking algorithm to compare data with VOT2018 short-term tracking data set;target tracking algorithm merged with handcraft and deep features,using VOT2018 long-term tracking data set perform data comparison.The experimental analysis tools use VOT-Toolkit to evaluate the short-term tracking algorithm's performance by accuracy,robustness and EAO comprehensive score,and the long-term tracking algorithm's performance by precision and recall.