On-Road Thermal Image Recognition by Efficient Edge Computing

This paper presents a method for thermal image recognition utilizing the YOLO v11 model deployed on Jetson Orin Nano. The system utilizes the Tensor RT cores to accelerate the recognition task, enhancing performance and power efficiency. Thermal image data for the recognition are from Teledyne FLIR Free ADAS (advanced driver assistance system) Thermal Dataset v2. The model is trained on the YOLO-v11 model, which achieves 71% precision scores on a small-size model and 59% of mAP50 (mean Average Precision at 50% Intersection over Union (IoU) for object detection) accuracy scores on a medium-size model. Results show Jetson Orin Nano can perform approximately 36.5 frames per second (fps) on a small-size model with 6 watts of total module power consumption or 1.9 watts on CPU and GPU cores. The YOLO v11 model performs approximately 15% better than the YOLO v8.