Mesh-Shrink: Real-Time Fast Moving Object Tracking with Sporadic Occlusion

Abstract

This paper proposes a new method for real-time object tracking in scenarios where the target moves fast but its appearance does not change quickly. By creating a mesh on the current frame, we set nodes of the mesh as candidate positions of the target. Several adjacent nodes which are more similar to the target constitute a new smaller search region, on which a new finer mesh is created. The approach iterates until certain conditions are satisfied. Finally, one of the nodes is identified as the target location. Unlike existing tracking methods, this approach achieves tracking of fast moving object in real time and is capable of recovering tracking when the target is missing due to full occlusion, or moving out of the image area and reappearing in the near future frames. The method does not require complicated computation and thus can be applied to the environment which permits only limited computing resources. The capabilities of the tracking based on our method are demonstrated by several image sequences.