ABSTRACT: This paper describes an occupant classification system based on Eigen shapes and support vector machines classification using 3D data. The occupant classification system is used to classify an occupant type into an adult, a child, a child seat, or an object. The inputs are depth images from a time-of-flight based depth camera. The images are first segmented from the background and normalized for threedimensional translations. The segmented images are projected into Eigen shapes that are constructed from a training set. The projections are used as feature vectors for a support vector machines classification algorithm. The Eigen features and support vector machines complement each other since the former is a linear transformation to reduce the dimensionality of the space, while the latter can deduce the non-linear aspects of these lower-dimensional features. Our experiments lead to several conclusions. First, we compare between Eigen shapes and knowledge based features, e.g. computer generated versus human generated features. The Eigen shapes provides better results compared to various combinations of knowledge-based features. The combination of Eigen shapes and knowledge-based features provide the best results, since both of these features capture different characteristics of images. The Comparison with intensity based analysis shows that the depth images are more suitable for this application. The support vector machines have also been shown to be superior to several other classification algorithms. The system provides more than 98 percent recognition rate with depth images as input. The failure cases include extreme deformations of the occupant that were not part of the training set.
Computer Vision and Pattern Recognition - Workshops, 2005. CVPR Workshops. IEEE Computer Society Conference on; 07/2005