Infrared (IR) thermal imaging target detection is essential for enabling robots to conduct all-weather inspections in field environments. This paper addresses two key challenges: the limited computing power of embedded systems onboard robots for real-time detection, and the low resolution of small targets in thermal imaging. To address these challenges, a lightweight detection algorithm based on an improved YOLOv7 framework is proposed. First, the network structure is pruned to enhance real-time performance on embedded devices. Subsequently, the backbone is optimized by integrating adaptive convolutional layers and a batchless normalization module. To improve small-target detection accuracy, multi-rate dilated 3D convolution is used to extract high-resolution scale-sequence features, which are subsequently fused via a Feature Pyramid Network (FPN). Finally, the SIoU-based position regression method is introduced in the prediction stage to improve regression speed and accuracy. Experimental validation on the NVIDIA Jetson Xavier NX platform using a nighttime thermal imaging dataset shows a 162% improvement in FPS, with only a 1.95% reduction in mAP compared to the original YOLOv7, meeting the requirements for real-time detection.