Target tracking method based on nonparametric online clustering

Abstract

Provided is a target tracking method based on nonparametric online clustering. The invention aims to track a specified target of interest in a given video online, predict the specific position and the bounding rectangular box of the target in a subsequent image frame, and extract the distillation characteristic of the target of interest based on a small deep network learned from a large deep network. The characteristic has strong appearance expression ability, discrimination ability and robustness. The invention further aims to perform nonparametric online clustering on the historical appearance of the target of interest, generate a target appearance template with historical information retained based on a weighted clustering center, and generate target spatial distribution using the target appearance template. Therefore, the accuracy of target tracking is increased, the risk of target drift is reduced, and the error rate is reduced.

Description

technical field [0001] The present invention mainly relates to the fields of video image processing, intelligent video monitoring, pattern recognition and the like, and specifically relates to a target tracking method based on non-parameter online clustering. Background technique [0002] Object tracking is one of the most basic and challenging core issues in the field of intelligent video surveillance. As a relatively low-level problem in computer vision algorithms, it has broad application prospects. For example, high-level intelligent analysis applications such as motion analysis and anomaly detection in the field of intelligent video surveillance are based on target tracking. Extracting the trajectory of moving objects in the scene can provide an effective basis for higher-level intelligent analysis in the monitoring scene. [0003] Most of the traditional target tracking methods initialize a certain target of interest in the first frame of a given video, and then trai...

AI Technical Summary

Problems solved by technology

This kind of method uses the online learning mode to continuously update the appearance model during the tracking process, ignoring the fact that the target of interest is easily affected by occlusion, disappearance-reappearance, and it is easy to cause model drift.

At the same time, since the learning process of the model relies heavily on the appearance model, when the target drifts, it is difficult to re-initialize the target model, resulting in a complete failure of tracking and poor tracking results

Images