Vulnerable road users’ detection with bionic-corrected multi-fisheye images and safety warning for ART

Autonomous-rail Rapid Transit (ART) as a new form of road transportation faces the challenge of colliding with other road users. However, a complete and affordable environmental perception system has not yet been implemented to ART. This paper proposed a vulnerable road users detection and safety warning method for ART with four modules. The image acquisition module captures the surrounding environment using eight fisheye cameras distributed around the ART. The image correction module applies a bionic longitude and latitude correction algorithm to rectify fisheye images, generating images closer to normal vision. The target detection module applies a SOD-YOLOv5n lightweight detection model and is tested on the SODA10M dataset and our dataset. The mAP has increased by 7.9% and 3.6% compared to the original YOLOv5n directly, and by 4.4% and 4.0% compared to the latest YOLOv8n, respectively. Additionally, it integrates a monocular fisheye camera-based ranging algorithm, providing distance between ART and detected objects. The speed control module adopts a hierarchical speed control strategy based on the PID algorithm. By reading the speed and distance information, different accelerations are output for speed control. We confirmed the feasibility of the proposed method through experimentation and simulation, thereby ensuring safety of road users.