Pillars detection for side viewed vehicles

Detecting the parts of a vehicle represents a topic of major interest for computer vision applications, especially for precrash systems. This paper proposes an artificial vision based technique that identifies the pillars of the lateral viewed cars. The novelty of the approach resides in the multi-layer classification scheme applied within the context of a stereo-based object detection system. From all the objects deetected by stereovision the side viewed cars are recognized, and for them the pillars are identified. This process of pillar identification is the result of a multi-layer classification that comprises: a rough object hypothesis refinement that selects only those objects that are likely to have one or two wheels, followed by an adaptive boosting classifier build using histograms of oriented gradient features. The boosted classifier realizes a fine selection of the wheel-based hypotheses and discriminates between side viewed vehicles and other objects in a traffic scene. The last step consists in the construction of a geometrical model of the pillars' region of interest for the identified side vehicles.