Tracking of rigid-bodies for autonomous surveillance

Abstract

For robotic surveillance systems, real-time knowledge of the motion of objects in the surrounding environment allows greater autonomy and interactivity. In some applications, orientation is of just as much interest as the position of an object. This paper presents a novel 3D model based method for tracking the full 3D pose of a rigid body. The proposed method projects a texture-mapped model of the target object back onto the camera's image plane at the target's current predicted pose. Optical-flow is used to correct the error between the predicted pose and the real pose. Finally, the pose in the next time-step is estimated using a motion predictor such as a Kalman filter (KF). The proposed tracking algorithm was tested using both synthetic and real video sequences of a 50/spl times/50/spl times/50 mm textured cube. This cube's pose was successfully tracked to within 2.5 mm positionally and 0.6/spl deg/ angularly. The cube was approximately 600-840 mm away from the camera.