A novel saliency detection framework for infrared thermal images

In this paper, we present a novel human detection method by devising a saliency framework on visual attention HOG features for infrared thermal imaging cameras. The proposed approach extends the saliency map by including the representation not only spatial features but also gaze distribution features. During thermal videos, the developed framework consists several computational stages: (a) the regions of interest areas are outlined based on saliency contrast; (b) the grids of HOG descriptor are selected to extract features in each image; (c) the training features are optimized by gaze visual attention map; (d) finally support vector machine algorithm is used to register positive human saliency model for trained classifiers. In order to validate our algorithm, we constructed a thermal infrared image database collected by real-time inspection system that contains labeled gaze attention map. The experimental results using this database demonstrated that our algorithm outperforms previous state-of-the-art methods for human detection tasks in thermal infrared images.