Feature-level image sequence fusion based on histograms of Oriented Gradients

For improving accuracy and robust property of human detection, fusion of image sequences captured from visible-thermal sensors is lucrative. Instead of performing it in pixel-level directly, we try to fuse object features by a novel image sequence fusion algorithm based on gradient feature (GFIF). The GFIF algorithm calculate gradients of input images to form a joint histograms of Oriented Gradient (HOG) descriptor, and these fused features are used to train linear support vector machine (SVM). In our experiment, a color-thermal surveillance dataset is adopted and several multi-resolution fusion algorithms are tested for comparing. By the Detection Error Tradeoff (DET) curves, the GFIF algorithm shows its superiority.