Pedestrian and Cyclist Object Detection Using Thermal and Dash Cameras in Different Weather Conditions

Ensuring the safety of cyclists and pedestrians has become imperative in our ever expanding urban centers. Despite advancements in vehicle safety technology, traditional cameras often fail in adverse weather and low-light conditions. This paper investigates the efficiency of integrating thermal cameras with dash cameras to enhance detection accuracy of vulnerable road users. We first collected and annotated datasets, comprising thermal and dash camera footage under various weather conditions. We then developed a deep learning object detection model using YOLOv8 and Roboflow. Separate models were trained for each camera, then fused to compensate for their individual limitations. It was observed that dash camera is prone to occlusions and varied lighting, whereas the thermal camera excels in low-light settings. The performance metrics for the thermal camera showed a total mAP50 of 0.92 and mAP50-95 of 0.52 for detecting both cyclists and pedestrians, reflecting a highly effective system with significant potential to improve road safety.