Discriminatively trained patch-based model for occupant classification

This study presents a vision-based occupant classification method which is essential for developing a system that can intelligently decide when to turn on airbags based on vehicle occupancy. To circumvent intra-class variance, this work considers the empty class as a reference and describes the occupant class by using appearance difference rather than the traditional methods of using appearance itself. Each class in this work is modelled using a set of representative parts called patches. Each patch is represented by a Gaussian distribution. This approach successfully alleviates the mis-classification problem resulting from severe lighting change which makes the image locally overexposed or underexposed. Instead of using maximum likelihood for patch selection and estimating the parameters of the proposed generative models, the proposed method discriminatively learns models through a boosting algorithm by minimising training error. Experimental results from many videos (approximately 1 630 000 frames) from a camera deployed on a moving platform demonstrate the effectiveness of the proposed approach.