Computer Vision-Based Dangerous Riding Behaviors Detection

This study addresses the need for detecting dangerous riding behaviors in the context of e-scooters. The research focuses on developing object detection and image classification models to identify dangerous rides, particularly instances where multiple people ride on a single e-scooter simultaneously. To accomplish this, an e-scooter dataset is created, and pre-trained models are trained on the dataset using transfer learning. A YOLOv4-based object detection model achieves 93.89% mAP and 90% F1 score in detecting e-scooters. However, a two-class YOLOv4-based object detection model, designed to differentiate normal and dangerous rides, performs less effectively with an overall mAP score of 56.54%. To overcome the limitations, several CNN-based image classification models are trained. Among them, the DenseNet-121-based model demonstrates the best performance, achieving a precision of 86%, recall of 76% and an F1 score of 81% when detecting dangerous rides. The developed models show good results and could contribute to improving safety measures and regulating the usage of e-scooters. By accurately identifying dangerous rides, potential risks from e-scooter usage can be mitigated