Adaptive Kalman Predictor-based Algorithms for Motion Estimation

Abstract

Estimating object motion from image sequences is a problem in robot vision. Many applications of motion-from-images, including robot manipulation, navigation and visual tracking, require algorithms that can estimate motion on-line and have strong noiseproof feature. Kalman predictors can meet the above requirements. Based on the affine model of a moving image, this paper discusses a Kalman prediction algorithm for estimating an object's 3d translational velocity from image sequences. At first, the current statistical model of a moving object is established. Then, according to the affine model of a moving image, a mathematical relation between the 3-D translational velocity of a moving object and its image motion paremeters is found. Finally, an adaptive one-step Kalman predictor is designed. To alleviate divergence of the predictor, initial states is estimated. Simulation results show that the performance of the adaptive one-step Kalman predictor is good.