Thermal-Based Pedestrian Detection Using Convolutional Neural Networks and Multi-Regions Dropout Technique

Abstract

Pedestrian detection is one of the most common research topics in the area of computer vision. It has a significant impact on the safety of autonomous driving and related surveillance applications. Recently, because of the rising of Convolutional Neural Networks (CNN) in deep learning techniques, object detection algorithms have made a significant breakthrough in their accuracy and robustness. However, most detection algorithms can only perform stably under the environments where lighting is sufficient. Nighttime pedestrian detection remains a challenging problem. In this paper, we present a thermal-based framework to solve the nighttime pedestrian detection problem, which utilizes the thermal camera and extends the Faster R-CNN method. In addition, a multi-scale model scheme is used to enrich the learning information of pedestrian features, and a feature region segmentation method is added to solve the occlusion issue. From the experimental results, it demonstrates that our proposed method achieves higher performance compared to current classic deep learning pedestrian detection methods.