TLD-based multi-view target tracking device and method

Abstract

The invention discloses a target tracking device combined with a multi-view camera and a multi-view target tracking method. The target tracking device comprises a video capturing unit and a video processing unit, wherein the video capturing unit is used for capturing video information of multiple original visual angles; the video processing unit is improved based on a TLD tracking algorithm so that the purpose of multi-view target tracking is achieved starting from a detection module and a learning module of the algorithm. For the detection module, if a total graph is scanned, much time is consumed. A movement of a target between two frames is a small movement, so that a detection region is shortened into a small region, namely, ROI, around the last-frame target box, and thus the tracking time is greatly shortened. The target tracking device further comprises a multi-view information fusion unit, wherein the multi-view information fusion unit fuses multi-view information, and thus the robustness of the tracking of the detection module is enhanced by using a learning mechanism, so that the tracking performance is improved. According to the target tracking device, by fusing multi-view target information, the target tracking performance is improved.

Description

technical field [0001] The invention relates to a target tracking algorithm under the condition of multi-view angles. It is mainly used in the field of sports posture analysis such as golf swing analysis, and can use multi-view information to improve tracking performance and efficiency. Background technique [0002] Moving target tracking is one of the core contents of machine vision research. It is very simple for humans to recognize moving objects from images. But there are various problems and obstacles for computers. The tracking of moving targets uses image processing and other related technologies to process, analyze and understand video sequence images. Find the location of the target of interest in the sequence image, and track the target to realize the behavior understanding of the moving target. Moving target tracking plays an important role in remote sensing and telemetry, industrial security, aerospace, intelligent robots, social security and other fields. T...

AI Technical Summary

Problems solved by technology

Monocular vision has a small amount of information, and only needs to process one image at a time. The calculation speed is relatively fast, but the image loses the three-dimensional information of the actual scene during the projection process, so it has irreparable defects.

When using methods based on monocular vision to detect and track moving targets, there are often problems such as target occlusion, light changes in surrounding scenes, and shadow interference. How to solve these problems has always been a difficult research point.

The literature "Magdi Mohamed. Decentralized Multiple Camera Multiple Object Tracking" (Wei Qu, Dan Schonfeld. IEEE International Conference on Multimedia and Expo, 2006) also uses distributed Bayesian inference to obtain the posterior spatiotemporal transfer of high-order joint state distributions. However, there are two main problems in the method proposed by the author: 1) When the posterior derivation of space-time recursion sets a variety of conditional independence assumptions between multiple observations and states, it is not universal, and it is not necessarily true in real tracking scenarios. Satisfied; 2) Multiple types of measurements are not used, and the calculation of observation and state likelihood is limited to the target position value, and the gray information of the target is not fused

Images