Pedestrian detection is a high-profile topic in computer vision, in part because it has great relevance to autonomous driving and intelligent surveillance applications. However, most pedestrian detection algorithms perform stably only during the daytime with sufficient illumination. At night, there is still room for improvement and many challenges exist. These challenges include occlusion caused by objects or crowds, and the problem of image background segmentation caused by environments with varying illumination. In this paper, we propose a nighttime thermal image pedestrian detection system, which can be viewed as an extension of the Faster region-based convolutional neural network (R-CNN) method. The proposed system can be used for static surveillance scenarios. First, a part model branch is proposed to realize the learning of partial pedestrian block features. Second, a segmentation branch is incorporated to strengthen the positioning of the pedestrian foreground. Finally, the branches are integrated through the fused loss function to enable joint training and optimization of the detection model. To evaluate the performance of the proposed model, we tested the system with several nighttime surveillance scenes. The experimental results show that the proposed method can effectively deal with the occlusion problem under challenging illumination environments and achieve performance levels superior to those of some state-of-the-art deep-learning pedestrian detection methods.