Object tracking method based on vision

Abstract

The invention discloses an object tracking method based on vision. The method comprises steps: each frame of image acquired by a mobile phone is used as input, a to-be-tracked target is selected by auser, and a multi-scale KCF tracker is operated; during the tracking process, the tracking confidence is calculated by a peak sidelobe ratio, and a tracking filter model is updated in the case of highconfidence; when the confidence is smaller than a certain threshold, the tracked object is proved to be blocked, and a YOLO (You only look once) detection process is operated at the time; and the YOLO detects multiple alternative targets, fine detection is carried out according to the already-learnt tracking filter model and the alternative targets, and the one with the maximum response value isthe tracked target. The method is accurate in target tracking and good in real-time performance.

Description

technical field [0001] The invention relates to an object tracking method, in particular to a vision-based object tracking method. Background technique [0002] Moving target tracking is one of the key research issues in the field of computer vision, and it has important applications in many occasions. One of its applications is the combination of three-axis stabilized PTZ to achieve tracking control. At present, the fast and effective tracking methods are the kernel correlation filter algorithm KCF, and the DSST algorithm with scale adaptation function added. KCF uses the circulant matrix in the area around the target to collect positive and negative samples, uses ridge regression to train the target detector, and uses the circulant matrix to be diagonalizable in the Fourier space to convert the matrix operation into a point multiplication operation, which greatly reduces the computational cost. It can improve the operation speed and make the algorithm meet the real-time r...

AI Technical Summary

Problems solved by technology

[0004] The disadvantages of the above approach are: 1. The model update process and the tracking result update are together, which cannot guarantee that the model drift can be reduced and the target can be kept up as much as possible; 2. When the tracking process is running, it is necessary to train a The detection classifier adds a lot of computational burden, which reduces the real-time performance of the algorithm; 3. The sliding window method is used in the detection process, which is inefficient and if the tracking model has drifted, it is easy to detect and track the wrong target

Images