Thermal-Based Pedestrian Detection Using Faster R-CNN and Region Decomposition Branch

Abstract

In this paper, we present an infrared thermal-based pedestrian detection method that can be applied in nighttime intelligent surveillance systems. Pedestrian detection plays an important role in computer vision and automation industry applications, which include video surveillance, automotive robot, and smart vehicles. Recently, the improvement in deep learning techniques, such as convolutional neural networks (CNNs), have significantly increased the accuracy of pedestrian detection. Normally, the optical cameras, e.g. charge-coupled device cameras, are the device used to capture images. However, considering the dark environments and the luminance variation issues, infrared thermal camera would be an effective alternative solution to nighttime pedestrian detection. On the other hand, occlusion is one of the commonest problems, which makes nighttime pedestrian detection more challenging. To address the abovementioned problems, this work presents a pedestrian detection framework which consists of Faster R-CNN and a region decomposition branch. The proposed region decomposition branch allows us to detect wider range of the pedestrian appearances including partial body poses and occlusions. From the experimental results, this work demonstrates better detection accuracy than the currently developed CNN-based detection method because of combining the multi-region features.