Reconstruction And Tracking Of3D Target Swarms And Dynamic Surfaces

Three-dimensional (3D) target swarms consist of a large number of individ-uals, these individuals move in3D space forming complex3D shapes and motion patterns. Target swarms are very common in nature, for example, many kind of birds fly in flocks and fishes may swim in schools, swarms can also be found in insects such as bees and grasshoppers. For many years, scientists have been very interested in the behaviors of these animals and have tried to find some common rules and contributing factors behind them. Unfortunately, the lack of effective techniques which can accurately measure the real3D trajectories of the individuals limits the quantitative study on these behaviors. Recent advances in high-speed cameras have made it possible to capture the dynamics of these groups at an acceptable resolution. But automatically recovering the3D trajectories from obtained2D video streams is still a challenging task.We present in this paper an automatic tracking system for3D target swarm tracking, multiple high-speed cameras capture videos of the moving targets, and the tracking algorithm of our system can recover3D trajectory of each individual automatically. Our tracking method makes full use of the weak visual cues and can handle occlusion and resist against distraction from image noise effectively. As demonstrated in the experiments, the proposed method is able to obtain accurate3D trajectories of the targets on both simulated data of different target density and real fruit fly data.3D deformable surface reconstruction and tracking is another important prob-lem in computer vision. The motion of many kinds of surfaces such as cloth and human face can exhibit complex non-rigidity, which can not be represented with several parameters, so building an effective model of these surfaces and estimat-ing corresponding parameters of the models are challenging tasks. We present in this paper a method which models the surface locally as a spatio-temporal planar patch with9parameters, and we also propose an region-growing mechanism which makes the parameters be correctly estimated. In order to keep the reconstructed3D models temporally consistent, we deform a3D mesh model with the recov-ered3D motion field. As demonstrated in the experiments, the proposed method is able to handle large motion and reconstruct accurate sub-pixel3D shape and motion.