Interpretation of image sequences by spatio-temporal analysis

The authors present a system designed to analyze any sequence of closely sampled image frames. They describe recent experiments conducted in a vision laboratory using spatial and temporal information. The importance of temporal information is illustrated by experiments on random dot images and a method to use this information by a best-first-search in the temporal domain is suggested. A more elegant algorithm for extracting motion information in more realistic images is then presented. Results on both synthetic and real image sequences are shown. The advantages and limitations of both approaches are highlighted. It is concluded that, in contrast to most previous approaches, the proposed system should be quite robust, as it is capable of handling occlusion and disocclusion, which are explicitly modeled