Computational Study Of Motion Perception

Motion analysis is a challenge for vision research. In recent years, a relatively complete theoretical system of motion perception has been established. As widely accepted in biological vision, there exist at least two distinct low-level subsystems analyzing motion:a first order system that responds to certain moving luminance patterns, and a second order system that responds to moving modulation of feature types, which are usually defined by contrast, spatial frequency, temporal frequency etc.From the perspective of computational analysis, we focus on the low-level motion perception mechanism, and establish the corresponding computational model to explore a new biologically inspired motion detection algorithm.Combination of psychology and neurophysiology outcomes, separable spatial and temporal filters in motion detector are designed, which collaborate to compute the motion based on the delay-and-comparison principle. Furthermore its compositional modules are thoroughly analyzed to reveal the functional connotation respectively.In consideration of computer applications, an elaborated version of the biological correlation model is proposed with different decision principle. The implementation is valided both on synthetic and real world image sequences. Preliminary experimental results show that the proposed detector with the Winner-Take-All decision has better robustness and anti-noise capability.In this paper, different types of second order motion are formulized and investigated in detail. We present that second order motions can be divided into three typical groups according to the modulation types:spatial modulate motion, temporal modulate motion and spatiotemporal modulate motion.Through the analysis of second order motion, a general nonlinear preprocessor, Texture Grabber, is proposed for detecting various types of motions. Experiments are conducted by correlation model preceded with the nonlinear processor. Preliminary analysis demonstrates that the proposed detector can capture effective information from different types of second order motions. The computational results are consistent with the previous suggestion that the second order motions are processed by nonlinear system.Finally, we discuss the relationship between first and second order motion information. Appropriate combination will obtain more reliable motion estimation, and the bio-plausible exploiture may bring some new advantage to computer vision practice.