Automatic objects behaviour recognition from compressed video domain

In this paper we present a system that, directly from compressed video domain, establishes a correspondence between objects in motion in a video scene and a concrete behaviour. This behaviour is expressed by using linguistic variables. Besides, with this fuzzy logic-based approach, the imprecision and vagueness of our primary source of information, MPEG motion vectors, is reduced. Proposed algorithms for segmentation and tracking are based on fuzzification of MPEG motion data. Once the tracking phase has finished, a linguistic model for each objective in the scene is generated and compared with each one of the behaviour models previously described in a linguistic manner. Finally, a practical application of this system for detection, tracking and behaviour analysis of vehicles in complex traffic scenes is presented.