Research On Long-term Visual Target Tracking Algorithm

Robust visual single target tracking algorithm is a research hotspot in the field of computer vision.In recent years,the research in this field has made some progress,and the proposed target tracking algorithms have achieved good results in tracking robustness and tracking speed.However,there are still many key technical problems to be solved for long-term target tracking under complex tracking conditions.This paper summarizes the achievements in the field of visual single target tracking both at home and abroad,and proposes two single target tracking algorithms based on traditional correlation filters and deep learning frameworks for long-term target tracking tasks in complex tracking scenarios.The main work of this paper is as follows:Firstly,a long-term visual target tracking algorithm based on dynamic correlation filter is proposed.In the aspect of object modeling,the object representation model is constructed by extracting Histogram of Oriented Gradient features and color features of the object and its local context,which makes the object representation model robust to geometric deformation and illumination changes.At the same time,by constructing the adaptive target response motion model in the target detection stage,the problem that traditional correlation filter based tracking algorithms are easy to drift in tracking is avoided to a certain extent,and the problems of fast target motion,motion blur and partial occlusion can be well solved by the algorithm.In the aspect of target rotation angle and scale estimation,a AKAZE key feature point matching method is proposed,and the active feature points register is set to dynamically register and update the key points after matching,so that the problem of target scale and rotation angle change which is difficult to solve in most target tracking algorithms at present is well solved.In the aspect of target redetection,a Compressed Support Vector Machine(CSVM)model is proposed for target redetection.According to the theory of compressed sensing,a compressed sensing matrix is designed,which satisfies the restricted isometry property(RIP),and the high-dimensional target feature vectors are projected into the low-dimensional feature space.Then,the compressed low-dimensional feature vectors are trained through grid search and cross-validation methods to obtain the accurate classifier model.When the tracking drift problem occurs or the target disappears from the image and suddenly appears,the Compressed Support Vector Machine model can accurately detect the target again,so that the algorithm can track the target for a long time under complex conditions.Finally,the proposed algorithm is evaluated on the international target tracking standard test platform(OTB100).The experimental results show that the proposed algorithm has good tracking accuracy and robustness,and can achieve an average tracking speed of 30.6 frames per second.Then,a long-term visual target deep tracking algorithm based on hierarchical extreme learning machine is proposed.In the aspect of object modeling,a new deep tracking framework including a classification network as well as a regression network is constructed to extract the object features which containing rich characterization information,so as to realize the accurate modeling of tracking object.In the classification network,the first five layers of the deep learning model called VGG16 after pre-training are used as the feature extraction layer.Particle filtering,feature mapping,affine transformation and other technologies are fused to generate candidate image region proposals with low noise and low redundancy.And finally these region proposals are sent to the binary classification network layer and corresponding outputs are obtained.When the classification network is updated,the parameters of the feature extraction layer are fixed,and only the last binary classification network layers are updated,thus ensuring the classification stability and greatly improving the speed of the algorithm.In the regression network,a lightweight hierarchical extreme learning machine model is proposed,which consists of an automatic sparse encoder based on extreme learning machine,a convolution feature extraction layer and a regression model based on online sequential extreme learning machine(OS-ELM)model.The first two parts are used as the feature extraction layer of the network,and the third part is used as the regression layer to realize the regression of the bounding box of the tracking target.At the same time,the tracking target bounding box of the whole network is output by designing a tracking trajectory optimization strategy,and a joint complementary network update mechanism is proposed to ensure the stability of the algorithm by analyzing the changes of model performance during online tracking and periodically updating network parameters in both long-term and short-term.Finally,the proposed algorithm is tested on the international target tracking standard test platform(OTB100 and VOT2016).The experimental results show that the proposed algorithm has good tracking accuracy,speed and robustness.