Methods and Approaches for Real-Time Hierarchical Motion Detection

The recent work on perception and measurement of visual motion has consistently advocated the use of a hierarchical representation and analysis. In most of the practical applications of motion perception it is absolutely necessary to be able to construct and process these hierarchical image representations in real-time. First, we discuss a simple scheme for coarse motion detection that highlights the capabilities of the PIPE image processor, showing its ability to work in both the spatial and temporal dimensions in real-time. Secondly, we show how this architecture can be used to build pyramid structures useful for motion detection, again emphasizing the real-time nature of the computations. Using the PIPE architecture, we have constructed a Pyramid of Oriented Edges (POE) which is a logical extension of Burt's pyramid and also a version of Mallat's pyramid. The results of these algorithms are available on a video tape to highlight their real-time performance on moving images. Third, we propose a new method using PIPE that will allow dense optic flow computation and which relates the intensity-correlation and spatia-temporal frequency based methods of determining optic flow. 'Current address: Robotics and Flexible Automation, Philips Labs. Briarcliff Manor. N. Y.