PCE-YOLO: One novel approach for infrared pedestrian detection

Abstract Infrared images are suitable for pedestrian detection in nighttime and low-light conditions. However, due to the difficulty in extracting pedestrian features from infrared images, existing object detection models struggle to maintain high detection accuracy and speed. To address this issue, an infrared pedestrian and rider dataset (IPRD) is introduced and an improved model based on YOLOv8s called PCE-YOLO is proposed. The improvements based on the original model include three parts: 1) designing a partial convolutional module (PCM) to extract deep features of pedestrians; 2) integrating the CBAM at the end of model’s backbone to enhance the model’s attention to important features of infrared images while reducing background interference; 3) incorporating EIOU Loss into model training to improve object localization accuracy. Experimental results show that the PCE-YOLO model achieves a high balance between accuracy and detection speed, which is significantly better than the existing detection models.