Multi-scale Correlation Filtering Tracking Based On Directional Space

Target tracking is a very practical computer technology.The target tracking algorithm detects the target position of the video sequence to be tracked by the computer device according to the initial information of the target,and then outputs the tracking result(target position and size).With the widespread use of target tracking technology in practical application scenarios,faster and better target tracking technology has become the focus of more and more scholars.The core of the target tracking algorithm is the extraction of target features and the matching of target templates.This paper focuses on the following aspects of the core problem.1.A correlation filtering target tracking algorithm based on Gabor filter is proposed.The method uses Gabor filter to obtain target features,extracts multiple features of target by selecting different scales and directions,and trains relevant filter parameters according to the feature groups extracted by Gabor filter,then calculates the corresponding correlation response diagrams of each group of features using the filter template,and gets the final target position by weighting the response diagrams of each filter.At the same time,according to whether the target occlusion occurs,we choose whether to update the filter parameters,so that the filter can adapt to the new target,so as to track the subsequent frames.The proposed method has a good tracking effect when the target has illumination change,shape change and fast translation.2.A multi-layer correlation filtering tracking method based on scattering transform is proposed.In this method,the scattering descriptor is used to build the feature model of the target,and the multi-layer correlation filter is constructed according to the scattering descriptor of the target.At the same time,according to the characteristics of multi-layer correlation filter,the weight adaptive updating method of each layer filter and the corresponding occlusion discrimination method are proposed,and the parameters of the kernel correlation filter model are updated.When the shape,illumination of the target change and occlusion occurs,the proposed method has good robustness.3.A multi-scale correlation filter tracking method based on the edge feature of direction space is proposed,which extracts the edge feature of direction space of the target area and combines it with the correlation filter.The weights of the filter are updated adaptively to determine the final target position through the weights of multiple correlation response graphs.Target feature descriptors of multiple scales are selected and the final target scale is determined by scale filter.When the shape,illumination and scale of the target change,the experiments show that the proposed method is robust and obtains the better results.