Occupant classification invariant to seat movement for smart airbag

This paper presents an occupant classification approach based on monocular vision for smart airbags that can decide to deploy or turn off intelligently. The main focus of this work different from those in the literature is on addressing the issue of the movement of car seat. The idea behind is to introduce the relation between the object of interest and scene inside the vehicle, namely, contextual information, for priming the seat configuration. As for circumventing the problem of lighting change as well as intra-class variance, we model each class by a set of representative parts called patches and describe the patch by using appearance difference rather than appearance itself in the tradition approaches. The selection of patches and the estimation of their parameters are achieved through a boosting algorithm by minimizing the loss of training error instead of using maximum likelihood (ML) strategy. Finally, we evaluate our proposed approach using a great amount of database collected from the camera deployed on a moving platform.