Human detection is a technology that detects pre-determined human shapes in the image and ignores everything else, which plays an irreplaceable role in video surveillance. However, modern person detectors have some inefficiencies in detecting pedestrians at night, and the accuracy rate is still insufficient. This paper presents a novel practical model for automatic real-time human detection at night-time. For this purpose, a new network architecture was proposed by improving the ting-yolov3 network for detecting pedestrians from TIR images based on the YOLO algorithm's tasks. The K-means clustering method clusters the image data, which contributes to obtaining excellent priority bounding-boxes. The proposed network was pre-trained on the original COCO dataset to obtain the initial weights. Through the comparison with the other three methods on the FLIR and DHU Night datasets showed that the proposed method performance was outperformed, in addition, to achieve a high score of accuracy (mAP%) in the TIR images. The method has a delay in detection time of 4.88ms. By improving the performance rates of human detection in TIR images, we expect this research to detect intruders in the night surveillance system.