<![CDATA[Autonomous vehicle perception systems predominantly use red-green-blue (RGB) cameras, which experience significant performance degradation under adverse lighting conditions, with detection accuracy declining by 35–45% at night and in inclement weather. This limitation poses substantial safety risks for practical autonomous vehicle deployment. This study developed a robust thermal image-based road segmentation system using a You Only Look Once (YOLO) v8 deep learning architecture integrated with trapezoid zone detection for real-time autonomous navigation capable of consistent all-weather operation. The methodology encompassed five stages, including collection and annotation of thermal images using a FLIR Boson camera across diverse road conditions, two-stage model training employing YOLOv8n-seg architecture with copy-paste augmentation on NVIDIA Jetson Orin AGX, development of a polynomial regression-based distance calibration system, implementation of a six-zone priority-based trapezoid navigation framework, and comprehensive system integration with performance evaluation. The experimental results demonstrated that the proposed system achieved road segmentation average precision (AP) of 98.3% at an intersection over union (IoU) threshold of 0.5, with an overall mean AP (mAP) of 83.3% across six object classes after the second training stage. The copy-paste augmentation strategy improved minority-class (people) detection by 14.0%. The distance calibration system attained an R² score of 1.0 within the measured range. The daytime-trained model successfully detected objects at nighttime without retraining, demonstrating the fundamental advantage of thermal imaging. The system processed at 29.8 frames per second with 18 W power consumption. These findings validate thermal imaging effectiveness for all-weather autonomous navigation and establish a framework for embedded autonomous vehicle deployment.]]>
