Moving Target Detection and Tracking Algorithm Based on Context Information

To improve the robustness of target tracking algorithms in a complex environment, this paper proposes the moving target detection and tracking algorithm based on context information and closed-loop learning. A context region is composed of the target region and its current neighboring background. For every frame that follows from a video stream, the long-term tracking task is principally decomposed into four parts of synchronous operation: tracking, detection, integration, and learning. First, the tracker obtains the posterior probability of the target location and estimates the target state over succeeding frames by exploiting the spatio-temporal local information. Meanwhile, the detector searches for the target in independent frames combining with the context information of tracker, and automatically reinitialize the tracker when it fails. Then, the integrator attains the best location of the target by merging the output results of tracker and detector together through an optimal strategy. Finally, the learning process is designed as the feedback and generates training samples to update the detector according to the results of tracker and detector. Experimentally, we evaluate the performance against several latest techniques on various benchmarks, and the results demonstrate that the proposed algorithm performs remarkably in terms of robustness and tracking accuracy.