Multi-channel feature based high speed related filtering object tracking method

Abstract

The invention provides a multi-channel feature based high speed related filtering object tracking method. The method includes (1) constructing an object feature model through weight fusion of three channels of an HOG feature, a CN feature and an angular point feature so as to perform long term stable tracking on the object; (2) utilizing the HOG feature to overcome a problem of influence on objectfeature extraction due to luminance change; (3) utilizing the CN feature to overcome a problem of texture information rapid change due to dimension change of the object; (4) utilizing the angular point feature to overcome a problem of object tracking failure due to partial shielding of the object; (5) converting each feature to frequency domain through Fourier transform for relevance calculationand convolution operation is converted to point multiplication operation so as to reduce the calculation volume and accelerate the calculation speed; (6) adopting the MPR (Maximal Peak Ratio) to judgewhether an weight coefficient of the object feature model requires update or not and adjusting the leading role of a specific feature channel in a specific environment factor in a self-adaptive manner; (7) utilizing online learning to select a fixed learning factor to update the weight coefficient and implementing and the object model updating process.

Description

technical field [0001] The present invention relates to a correlation filtering target tracking method based on multi-feature fusion. Aiming at the interference problem of target tracking in different scenarios, multiple feature fusion is used to determine the location of the target, and the tracking process of the target in the video sequence is completed. . Background technique [0002] Object tracking is an important research direction in the field of computer intelligent vision. It uses the image sequence generated by optical lens and other hardware facilities to process and analyze, so as to obtain the specific coordinate information of the target in the video sequence. With the in-depth research in the field of target tracking, many excellent target tracking methods have emerged, such as inter-frame difference method, optical flow method and background model subtraction method, in addition to the tracking learning detection algorithm (TLD) proposed by Kalal et al. ),...

AI Technical Summary

Problems solved by technology

[0003] However, there are many factors in the actual application environment, which lead to problems such as illumination changes, partial or complete occlusion, non-rigid deformation and size transformation in the target video, which seriously affect the accuracy of traditional target tracking algorithms.

Illumination changes mainly change the gray value of each pixel in the target area, resulting in the failure of the judgment of certain features; partial or complete occlusion is often due to obstacles such as buildings in the video environment, which destroys the integrity of the target outline; Rigid deformation and size change will cause the information of the target contour or texture to be seriously disturbed in the gradient direction

Although many researchers have proposed many tracking algorithms based on various mathematical theories, it is difficult to achieve robustness to various environmental factors, which makes it difficult to overcome the influence of various environmental factors in the target tracking algorithm. more practical

The judgment criterion used by traditional correlation filter target tracking algorithms is generally a single feature. Common features such as the Histogram of Oriented Gradient (HOG) feature are the first-order derivatives of the image, which often affect the illumination changes of the environment. It has good robustness, but it is greatly affected by factors such as target size changes and shape changes, and the color name (Color Name, CN) feature is the zero-order feature vector of the image, which is very good for size and other appearance changes. Invariant, but easily disturbed by changes in lighting

In addition, there are corner features, which are prominent and representative points in the image, which allow the target to match the positions that appear in the two frames before and after the target is partially occluded, thereby reducing the cost of extracting certain features of the target. Necessary area integrity requirements, but when similar targets appear at the same time, the extracted corner features are easily disturbed, resulting in target tracking failure

Images