Abstract
We present a three–level architecture for detection and tracking of independently moving objects (IMOs) in sequences recorded from a moving vehicle. At the first stage, image pixels with an optical flow that is not entirely induced by the car’s motion are detected by combining dense optical flow, egomotion extracted from this optical flow, and dense stereo. These pixels are segmented and an attention mechanism is used to process them at finer resolution at the second level making use of sparse 2D and 3D edge descriptors. Based on the rich and precise information on the second level, the full rigid motion for the environment and for each IMO is computed. This motion information is then used for tracking, filtering and the building of a 3D model of the street structure as well as the IMO. This multi-level architecture allows us to combine the strength of both dense and sparse processing methods in terms of precision and computational complexity, and to dedicate more processing capacity to the important parts of the scene (the IMOs).