Accurate seat belt detection in road surveillance images based on CNN and SVM

Seat belt detection in intelligent transportation systems is an important research area, but the current algorithms for such systems are not very well developed. Existing methods are mostly based on edge detection and the Hough transform. However, there are many kinds of vehicles and background environments, which produce many possible edges; thus, these methods often produce false positives. We therefore propose a seat belt detection algorithm for complex road backgrounds based on multi-scale feature extraction using deep learning. We first extract multi-scale features from the regions of the labeled vehicle, windshield, and seat belt to train the detection models using convolution neural network (CNN). Then the coarse candidates of the vehicle, windshield, and seat belt in the test image are detected. For the accurate detection results, a post-processing is employed by using the detection scores as well as the relative positions of these vehicle components to train a classification model through support vector machine (SVM). Finally, we perform a fine mapping and identification process using this classification model on the seat belt region. This method performed well when applied to a database of images collected by road surveillance cameras.