Multi - View Large - Scale Particle Swarm Motion Trajectory Reconstruction

Animal swarms behavior are spectacular and magical phenomenon in nature. Besides bird flocks, fish schools and insect swarms, there are some swarms we are difficult to observe with the naked eyes, such as microorganisms and bacteria group. Scientists from various fields have been attracted by these phenomena and made a variety of methods to analyze it from different perspective. Acquiring motion data of individual is essential to accurately understanding the large-scale swarms’movement. However, the traditional biological research methods can only get motion data of a single or a small amount of individuals. It is difficult to effectively simulate the movement behavior of real groups. With the rapid development of the computer technology, using imaging equipment with computer vision methods makes it possible to get the trajectories of movement group without interfering and is attracted to more and more researchers. It is also a research focus in computer vision.The general method of acquiring the three-dimensional trajectories of swarms by computer vision technology is capturing a series of images of swarms from different views using accurate synchronization cameras, and then reconstructing the trajectory of target according to projection geometric relationship. However, it is difficult to use object detection algorithm with image texture features to match hundreds of targets due to similar shape and texture of individuals. Besides, the particle may be blocked by other individuals frequently or temporarily leaves because of perspective and vision. It is a great challenge to acquiring trajectories in a long period. Furthermore, it needs several hours to complete the calculation, because the computation rapid grows with the number of individuals. This paper proposed a method, which can acquire the trajectories of individuals in large particle swarms automatically, accurately and fast.In our method, a third imaging device, which obtains the videos of group from different view, is used as verification view. The verification view can significantly reduce the stereo matching ambiguity by filtering mismatched pair according to additional epipolar constraint. Consequently, the computation can be reduced with it. The proposed method employs optimal assignment with state prediction and candidate filtering to establish temporal association of particles. In state prediction, Kalman filter based on three-dimensional coordinate is used to predict the position of the target, aiming to reduce the trajectory fragment. Experimental results show that proposed method can accurately obtain the three-dimensional trajectory of the simulated particle swarm within tens of seconds, and can get hundreds of three-dimensional trajectories of fruit flies.Quantitative analysis of group behavior by acquired trajectories of targets is also one of focuses of this paper.