Multiple Object Tracking for Similar, Monotonic Targets

As a vital and extremely hot branch of computer vision, multiple object tracking is widely used in security monitoring and biological behavior research. At present, the majority of the multiple object tracking algorithms are composed of two stages: detection and data association. There is often mutual interference between targets in each frame, which causes great trouble to object detection and correlation. Especially for the microscopic particles, detection and tracking always is a tricky business because of its similar appearance and irregular movement. In this paper, a multiple object tracking algorithm based on two-step detection is proposed for microscopic non-fluorescent labeled object with simple structure, monotonous features, even no obvious features for learning. The specific operation is: first threshold segment each image in video to extract targets from the background, and perform morphological operations on the obtained graph to separate the slightly adhesion targets to avoid incorrect tracking, then estimate object states by Kalman Filter and use Hungarian algorithm for data association, definitively achieve the goal of multiple object tracking. Experiments show that our algorithm is better than the general method, meeting the tracking task of multiple non-fluorescent labeled particles commendably.