Seat detection in a car for a smart airbag application

We present a method to detect the seat and head-rest of the by-passenger, as a part of a smart airbag system. The recognition of the seat and head-rest is useful for the purpose of background subtraction, as well as for assisting head-tracking and occupant classification. We use a multi-resolution probabilistic generalized Hough transform (GHT). We present experimental results for the detection, as well as an error analysis. Our experiments were performed using an imperfect set of models on close-range images with low dynamic range and under sever occlusions. Nevertheless, we have found that one needs to consider only the best 11 hypotheses of the GHT to ensure recognition. Moreover, when at least 25% of the seat contour is not occluded, only two hypotheses are needed on the average. The results show that the head-rest is a more robust clue than the seat. Finally, we discuss how to extend our work and possible uses in the context of occupant detection and classification.