TLD target tracking method and device based on Mean-shift model optimization

Abstract

The invention discloses a TLD target tracking method and device based on Mean-shift model optimization. The method comprises that The method comprises that a tracking target required by a first frame is marked, the gray color of a selected area is divided uniformly to obtain gray-scale histograms at the same interval, and a target model is obtained; when a tth frame is tracked, according to the target center position f0 of a (t-1)th frame, the coordinate f of a candidate target center position is obtained by taking f0 as the center of a searching window, and a candidate model is obtained; the similarity between the target model and the candidate model is calculated to obtain a similarity result; and according to the similarity result, iterative computation is carried out on candidate of the target model to obtain a new target area. According to the invention, the target area obtained in a Mean-shift algorithm is fused with the target area obtained by a TLD model tracker, the tracking model has a good tracking effect during modal rotation and partial shielding, the tracking effect of the tracker can be improved, the anti-shielding capability of the model is improved, and the capability of identifying similar targets is improved.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a TLD target tracking method optimized based on a Mean-shift model and a device thereof. Background technique [0002] Due to the model used by the tracker and detector in the core of the standard TLD model, the overall TLD model will also have obvious defects in the core. Among them, the tracker is the biggest problem. Since the optical flow method is used to track the target, the model is more sensitive to illumination. When the shape of the tracked object is rotated, the tracker will gradually move away from the tracked object, and the tracker will be re-initialized only when the object appears again in the learned shape. And when there are objects with high similarity in the background area, the optical flow method is easy to predict errors, resulting in the detector training samples containing wrong positive sample objects. Therefore, if there are more similar obj...

AI Technical Summary

Problems solved by technology

And when there are objects with high similarity in the background area, the optical flow method is easy to predict errors, resulting in the detector training samples containing wrong positive sample objects

Therefore, if there are more similar objects in the video for a long time, the positive samples of the detector will contain more error information, and then the detector will also cause initialization errors due to accumulated errors during the long-term tracking process.

Images