Segmentation, classification, and tracking of humans for smart airbag applications

There has been considerable attention paid to developing ‘smart’ airbags that can determine, not only, if they should be deployed in a crash event, but also with what force they should be deployed. Information on the size and type of front passenger seat occupant is used to determine the safe level of force with which to deploy the airbag. To date vision systems have been successfully applied to relatively controlled environments, such as for manufacturing, or they have been used in uncontrolled environments, such as for surveillance, where there is a human in the loop to monitor the performance of the system. In this thesis we have developed a computer vision-based approach to airbag suppression that attempts to provide the robustness required for an autonomous system fielded in a relatively uncontrolled environment. It addresses a very difficult real-world application, in which computer vision had not previously been applied, to simultaneously perform real-time human recognition and tracking. The specific contributions to pattern recognition include the development of a new filter-based feature selection algorithm based on robust statistical measures of discriminability and feature correlation. The algorithm performs as well as, or better than, other filter or wrapper methods, and is considerably faster. We have also defined a contextual processing algorithm that uses a continuous stream of classifications and the theory of evidential reasoning. This stream of results is integrated using the Dempster-Shafer rules of belief revision, with the available image information. Another contribution is the development of a unique wrapper-based image segmentation algorithm. We have adopted a paradigm where an image is initially region labeled and, then, using proven feature selection methods, we group these regions based on our knowledge of the desired object being segmented. This algorithm is shown to provide segmentations as accurate as human hand segmentation in many cases. There are also two contributions to the area of human motion tracking. The first is the definition of an information theoretic motion segmentation algorithm. This approach appears immune to illumination effects, and it dramatically changes the way we perceive image motion through information flow rather than optical flow. Lastly, we have developed an integrated motion and shape tracking system based on interacting multiple models (IMM) Kalman filtering. The approach is superior to HMM-based tracking systems, since it intelligently blends the individual dynamics states. Also, the system has been shown to be able to react to high-speed motion events, such as a pre-crash braking event. As part of the tracking system, we have also defined a new mechanism for inferring the 3-dimensional pose of the occupant based on the gross changes in their shape during motion, called shape from deformation.