This paper presents a high-speed and robust dual-band infrared thermal camera based on an ARM CPU. It is composed of a low-resolution long-wavelength infrared detector, a digital temperature and humidity sensor, and a CMOS sensor. In view of the phenomenon of large contrast between face and background in thermal infrared image, this paper we search for a suitable accuracy-latency tradeoff for thermal face detection and propose a tiny-lightweight detector named YOLO-Fastest-IR. Four different scale YOLO-Fastest-IR0 to IR3 thermal infrared face detectors based on YOLO-Fastest are designed. To train and test four tiny-lightweight models, a multi-user low-resolution thermal face database (RGBT-MLTF) is collected, and the four networks are trained. Experiments reveal that the lightweight convolutional neural network can also perform well in the thermal infrared face detection task. And the algorithm is superior to the existing face detection algorithms in positioning accuracy and speed, which is more suitable for deployment in mobile platforms or embedded devices. After obtaining the region of interest in the infrared image (IR), the RGB camera is guided by the results of thermal infrared face detection, to realize the fine positioning of RGB face. The experimental results show that YOLO-Fastest-IR has a frame rate of 92.9 FPS on a Raspberry Pi 4B and can successfully locate 97.4% of the face in the RGBT-MLTF test set. The integration of infrared temperature measurement system with low cost, strong robustness and high real-time performance was ultimately achieved, the temperature measurement accuracy can reach 0.3 degrees Celsius.