Real-time tracking method of nonspecific target based on partitioning

Abstract

The invention relates to a real-time tracking method of a nonspecific target based on partitioning, comprising three steps of classifier updating, target detection and weight updating. In the method, the target region is divided into multiple blocks; a classifier is used for maintaining each block, and updating is conducted frame by frame; the detection result of each classifier is comprehensively considered to determine the position of the target in the new video frame. In the method, an automatic weight updating mechanism is designed, so as to enable the blocks which are relatively stable to have greater decision-making power over the judgment of results, thus reducing the influence of various interferences; and the tracking performance is better than multiple international published algorithms recently. In the method, changes of appearances of objects caused by various interferences can be captured and accurate tracking can be conducted; the method has universality on target objects in various shapes and types; the calculation has low complexity and can be processed at real time. The invention has wide application prospect in various occasions needing tracking techniques, such as video monitoring, automatic driving, man-machine interaction, intelligent traffic, robot, airborne early warning and the like.

Description

technical field [0001] The invention relates to a method for tracking a target in a video, in particular to a block-based non-specific target real-time tracking method. Background technique [0002] Object tracking is an important work in the field of computer vision and automation, and has a very wide range of applications in military and civilian applications. Such as human-computer interaction, intelligent transportation, security monitoring, robots, air early warning, etc. These applications have put forward higher demands and requirements for automatic real-time target tracking. [0003] At present, there are many methods for video object tracking, such as those based on contour or template matching, based on filters, based on classification, and so on. The more prominent ones are the mean-shift tracking algorithm proposed by Dorin Comaniciu in the last two years, the fragment-based tracking method (Frag Tracker) proposed by AmitAdam, the (supervised / semi-supervised) ...

AI Technical Summary

Problems solved by technology

The first two methods are weak in tracking the target with significant changes in target motion and appearance. The online Adaboost method can better adapt to the change in the appearance of the target, but it is prone to drift. If the tracking time is longer, the target may be lost. The processing ability is weak. Compared with the online Adaboost method, the fourth method enhances the adaptive ability to the appearance of the target, and also has a certain occlusion processing ability, but the computational complexity is high and the real-time performance is low.

In addition, these methods track the target as a whole, and cannot distinguish different parts of the target. Local appearance changes caused by factors such as posture and occlusion will have a significant adverse effect on the tracking results.

Bo Wu et al. proposed a segmented pedestrian tracking method in 2007. This method is dedicated to the tracking of pedestrians. Its blocks are structured head, torso, etc. with semantic probabilities, which require pre-trained body parts. Part detectors are not universal for generalized targets

Images