To address the low precision and speed of existing pedestrian detection methods for thermal infrared images, a real-time pedestrian detection method based on a weak saliency map is herein proposed. The proposed method comprises two improved networks, namely, SD-LFFD and SF-LFFD, which use lightweight LFFD as the basic network. First, the thermal infrared image is input into the SD-LFFD to produce the preliminary pedestrian detection results and a weak saliency map indicating the pedestrian regions. Then, the weak saliency map and the original thermal infrared image are combined to highlight the potential pedestrian regions and generate new results using the SF-LFFD. Finally, the pedestrian detection results obtained by the two improved networks are integrated to obtain the final results. The experimental results on the CVC-09 and CVC-14 datasets indicate that the proposed method significantly improves the average precision (AP) of pedestrian detection compared with that of existing lightweight neural networks, and that it achieves real-time detection with limited hardware resources.