Research Of Object Tracking And Trajectory Recognition In Complex Scenes

The increasing demands in every area promote the development of intelligent video analysis technology. As one of the hottest research in computer vision, tracking and trajectory recognition cause more and more attention. However, since its high requirement of robustness, real-time and stability, there are still many problems to be solved in complex scenes.First of all, we research some classical algorithms of motion detection. By using ViBe(visual background extraction) with median fitter, the foreground can be detected without "Ghost". The pixel-level method, based sampling model, runs fast and robustly in real-time system. The foregrounds could be segmented more completely after erosion and dilation.To tracking the objects accurately and real-timely in complex scene, we firstly introduce compress tracking, which could reduce the image features to very low dimensions, and keep the features information almost. Given the problem of scale variation in the scene, the objects' multi-scale Haar-like features are extracted, and the searching should be extended into multi-dimension space. So the optimized searching strategy with PSO(Particle Swarm Optimization) is proposed. The result of some experiments demonstrate that the objects can be tracked stably even though partial occlusion, shape variation, scale variation, and the illumination changes.After detecting and tracking objects in complex scene, the trajectories could be obtained roughly. However, the length and noises problems seem inevitable. Thus, we cluster the trajectories via K-means and improved Hausdorff distance after pre-treating the trajectories feature. Finally, based on the results of cluster, we train the HMM(Hidden Markov Model) for every kind of trajectories. The patterns of trajectories can be recognized in complex scenes.Last but not least, a simple real-time detection and tracking system was designed and implemented, the PSO-CT algorithm performs well in real scene.