Marker-Less Vision-Based Detection Of Improper Seat Belt Routing

Seat belt routing classification is of crucial importance to the automotive industry. It determines whether an occupant's seat belt is not only buckled but also properly routed through the body, making sure that the occupant is truly protected in case of an accident. Current technology, such as the buckle Hall effect sensor, only addresses the problem of buckling detection, but no reliable solution exists for the proper routing scenario. In this work we compiled a database of more than 34,000 images each from two different camera setups and trained deep neural networks to detect improper seat belt routing. The results show that the best camera setup can achieve an accuracy of 0.977 on a coarse (four-class) wrong routing classification task and 0.924 on a fine-grained (seven class) classification task, which may be a step toward reducing injuries.