Real-time object detection and tracking on a moving camera platform

This paper presents a real-time tracking system to detect and track multiple moving objects on a controlled pan-tilt camera platform. In order to describe the relationship between the targets and camera in this tracking system, the input/output hidden Markov model (HMM) is applied here in the well-defined spherical camera coordinate. Since the detection and tracking for different targets are performed at the same time on a moving camera platform, the detection and tracking processes must be fast and effective. A hybrid detection algorithm which combines the target's color and optical flow information is proposed here. A two layer tracking architecture is then utilized for tracking the detected target. The bottom level utilizes the Kanade-Lucas-Tomasi (KLT) feature point tracker which identifies the local point correspondence across image frames. The particle filter at top level, which maintains the relation between target and feature points, estimates the tracked target state. The overall performance has been validated in the experiments.