Real Time Multi-Object Detection And Tracking Based On Feature Point Classification

Multi-Object detection and tracking is an important problem in Computer Vision, and has various applications in the security surveillance area, such as intelligent transport system. The classic algorithms have kinds of limitations. For instance, background subtraction based methods, are very difficult to handle problems of occlusions, illumination change, and camera shaking. Approaches based on feature tracking and grouping is usually not stable enough due to different sizes of objects. Other approaches which based on appearance matching, is usually time consuming and may not achieve real time performance. In this paper, we propose a fast object detection and tracking method based on feature point classification, which is independent of background model. Our method consists of two parts:1) in offline module, we divide the object into several parts according to feature point locations, and extract each part's feature to train a classifier; 2) in the online module, we first get the feature points and use trained classifier to decide which part the feature points belong to. Then we employ a discrete voting scheme to detect target objects in a very fast speed. Finally, after establishing the correspondence between feature points and objects, we track objects basing on feature point tracking. Our method is robust to partial occlusion since we use a part-based strategy. Experiments show that our method is very fast and effective. We also estimate the speed of object and segment the object basing on our method, both show good results.