Mass Group Movement Based On Computer Vision 3 D Trajectory Capture And Performance Optimization

Phenomena of large scale moving group are common in nature. Examples include the group motion of birds or other animals, rain drops and snowflakes, microscopic individuals like cells and bacteria, and etc. In the fields like collective behavior and fluid mechanics, researchers pay great attention to the motion of the individuals in the group in order to explore the underlying principles of their collective motion. As a result, there is a demand to develop automatic three-dimension trajectory measurement systems to measure each individual in the group. Among methods utilizing GPS systems and sonar technology, computer vision systems using image acquisition devices and image analysis algorithms, are becoming more widely used for their flexibility, feasibility and accuracy with the rapid development of computer vision technology.Stereo vision system is one of the best known computer vision systems to measure3D trajectories of moving individuals. These systems utilize multiple viewpoints like video cameras to capture images sequences of the group and acquire3D positions via triangulation. The major problems which stereo vision systems need to solve include:3D reconstruction, which produce stereo matching and reconstruct3D positions, and tracking, which associate observations at different time steps to a common real world individual. In this paper, stereo vision is categorized into:first tracking system (FTS), first reconstructing system (FRS), and parallel system (PS). To make full use of all available constraints, FTS and PS adopt motion clues from temporal tracking to solve the stereo matching problem.Nevertheless, facing large scale moving groups, which appear to be dense and similar, stereo vision systems suffer from problems like stereo matching ambiguity especially under binocular settings. Multiple stereo correspondences exist for a single projection and the true correspondence cannot be determined. Although motion clues are applied to solve the problem, the efficiency of disambiguation via motion is limited when focused on ambiguous situations like group translation and group rotation. Motivated to address the challenges, a new approach of viewpoint placement optimization is introduced to help the disambiguation of stereo matching ambiguity via motion clues. Normalized ambiguity rate (NAR) and ambiguity decreasing rate (ADR) is defined to quantitatively analyze the stereo ambiguity level, and on the basis of both theoretical analysis and experimental results, the optimal viewpoint placement is derived for measuring group translation and rotation so that the stereo-matching ambiguity decreases as quickly as possible. It is also demonstrated that the optimal viewpoint placement can significantly improve the robustness and efficiency of existing binocular methods that utilize motion clues to disambiguate.