Abstract:
Urgent attention should be paid to the safety of pedestrians at crosswalks in the face of growing traffic density and the lack of effective passive road infrastructure. This work proposes, develops and tests an edge-computed Smart Crosswalk System using real-time computer vision to adaptively control traffic intersections. The proposed system design is based on an NVIDIA Jetson Nano 4GB in combination with a YOLOv5n deep learning algorithm, with an optimised TensorRT implementation for edge computing. The computer vision system scans a Region of Interest (ROI) and identifies people and vehicles. When a perceived threat is detected, the unit activates an ESP32 microcontroller, enabling a localised warning system comprising of P5 SMD LED display panels, warning flashes and an automated alarm system. Experiments and evaluation demonstrate that the enhanced vision system is capable of maintaining a consistent real-time inference rate of 30 Frames Per Second (FPS). The object detection model obtained a mAP@0.5 of 90.9 percent with a class-wise precision of 91.1 percent for people and 90.8 percent for vehicles. The highest F1-score of 0.87 confirms the accuracy of the system in predicting the system’s performance. The results show that this system is an intelligent safety system that is fast, online, and economical which can be incorporated into smart city designs.
