Object tracking and classification system using agent search

Many methods for video surveillance have been developed. Robust tracking and high-performance classification are required for secure surveillance. In this paper, we propose a new recognition system that can track moving objects such as pedestrian, and classify them using a single camera in an open space parking. The proposed system can 1)perform robust object tracking and classification over occlusion and crossing; 2)look a local region of object image; and 3)integrate all processes into time series data flow. For object tracking, we developed a new agent tracking algorithm. A number of agents are generated for each object, and independently search and move to the future position by looking a local region's feature of their objects. Then they agents get fitness values, and the object ID of its local region is updated. For object classification, we forge a strong classifier from weak classifiers using AdaBoost. In practice, we recorded some scenes in an outside parking using a video camera, and tried to track objects and classified them into "person" or "vehicle". As a result, we achieved over 97% for tracking success rate, and over 87% for classification success rate.