Thermal Infrared Human Recognition Using CNN-based Abstraction of Multi-scale Monogenic Signal Representation and SVM

A method for highly accurate human recognition in thermal infrared images is proposed. Firstly, the images are segmented and post-processed, so that the regions of interest (ROIs), which may possibly contain human subjects, are available. Then, multi-scale monogenic signal representation (MMSR) of the ROIs is computed in order to reach intermediately discriminative representation of potential human bodies. Next, each MMSR component is abstracted via a respective CNN, and therefore the feature that is ready for classification is formulated by aggregating the full-connected layer output of each CNN. At last, based on the available feature, a trained support vector machine (SVM) classifies a ROI as a person or not a person. Experimental results shows that the proposed method has the advantages in recognition accuracy, efficiency, and promotion of feature interpretability.