Efficient seat belt detection in a vehicle surveillance application

In this paper, we propose a novel approach that determines whether the seat belts in the vehicle are belted or unbelted. It is a challenging problem because of some practical constraints including low quality images due to severe illumination conditions, view variation, complex background, etc. In order to alleviate these problems, our proposed approach can jointly train multi-detectors. It keeps the score map output by a detector and uses it as contextual information to support the decision at the next stage. Haar-like features and Histograms of Oriented Gradients (Hog) features are combined together to form more powerful image representations. Through AdaBoost learning, the most discriminative feature set is selected automatically. To verify the effectiveness of the proposed method, we evaluate our seat belt detection algorithm on six surveillance videos. Experimental results convincingly demonstrate robustness and efficiency of our system.