Research On Object Tracking Algorithm Based On Kernelized Correlation Filter And Its Implementation On Embedded Platform

As an research hotspot in the field of computer vision,visual object tracking technology is widely used in both civilian and military fields.After years of research and development,researchers put forward many excellent classical object tracking algorithms.However,in the face of challenges in the practical applications,such as illumination variations,object deformation,scale variation,object occlusion and so on,all the problems are still difficult to be solved at once by using the existing visual object tracking algorithms,which is still a very challenging task in the field of computer vision.In the visual object tracking algorithm,object tracking algorithm,based on kernelized correlation filters with real-time and high tracking precision is more suitable for implementation in embedded platform with limited hardware resources.How to improve object tracking algorithm will be discussed in this paper,which is based on the kernelized correlation filters.Firstly,an adaptive kernelized correlation filter tracking algorithm based on multifeature fusion is proposed in this paper.In order to solve the problem of using single feature to representational target,the kernel correlation filter with features of directional gradient histogram(HOG)and color name(CN)is trained to track the object and the response map of different features is fused by the confidence function of the response map at the decision level to improve the accuracy of the tracker;in order to solve the problem of using fixed update rate to update the model,the confidence of multi-feature response map is used to judge whether the object is occluded or not,to change the learning rate adaptively to avoid model drift;in order to solve the problem of using fixed target size,the scale detection module is introduced to improve the accuracy of the algorithm.Experiments indicate that problems,such as object occlusion,fast motion,scale variation and so on,can be effectively solved by the improved algorithm.Secondly,a kernelized correlation filter tracking algorithm with multi-hierarchical deep features is proposed in this paper.In order to solve the problem of limited representation ability of hand-craft features to objects,the convolution neural network is used to extract multi-hierarchical deep features,which are combined with hand-craft features and shallow features for object tracking,and then the response map is fused at the decision level for the improvement of the algorithm accuracy;meanwhile,the deep feature is used to train the kernel correlation filter to track the object,which is used to generate the weight map to constraint the fused response map generated before,and then the algorithm robustness can be improved.Experiments indicate that the tracking performance of the algorithm can be further improved by comparing with algorithms only with hand-craft features and combining algorithms with multi-hierarchical deep featuresFinally,considering the requirement of real-time in practical application,algorithm proposed in this paper,is transplanted to the ARM development board platform and optimized by multi-thread based on characteristics of embedded platform.Test results of system show that,under the premise of real-time,algorithm,can be run on the embedded platform under the limited hardware resources,has good practicability and expansibility.