Performance Analysis of YOLOv7 Training Methods in Autonomous Vehicles: A Study on the Hanwha IR Thermal Image Dataset

Autonomous vehicles (AVs) must accurately detect and identify objects in varying environmental conditions, which can be challenging in low-light or poor weather scenarios. Thermal imaging provides an alternative to traditional RGB vision, enhancing object detection in such conditions. This paper explores the effectiveness of two training approaches-full training and transfer learningusing the YOLOv7 model on the Hanwha IR Thermal Image Dataset for object detection in autonomous driving applications. The study aims to evaluate and compare the accuracy, computational efficiency, and practicality of each training method for thermal-based object detection, providing insights into the best approach for real-world applications in AVs. Results indicate that transfer learning is more efficient in training time and computational resource while achieving comparable accuracy with a mAP of $49.7 \%$ to full training with a mAP of $46.7 \%$, making it a viable option for real-time applications in autonomous driving systems.