YOLO-DC: Robust electric bike detection and tracking in critical scenarios

With the rapid growth of the global economy, transportation safety challenges have become increasingly prominent, resulting in substantial losses to life and property. Electric bikes (E-bikes) exhibit relatively high speeds and poor dynamic stability, leading to elevated driving risks, making accurate real-time E-bike detection a critical requirement for road safety and intelligent driving systems. Despite the unique significance of dedicated E-bike detection in Intelligent Transportation Systems (ITS), research on E-bike detection as a distinct category remains limited, compounded by a lack of open, complex-environment E-bike datasets. To address these gaps, we design a novel YOLO-DC model to improve the accuracy and real-time performance of E-bike detection in complex traffic environments. We enhance the YOLOv8s model with three targeted modifications: integrating Deformable Convolutional Network-v2 (DCNv2) into the backbone for improved spatial adaptability, embedding the Convolutional Block Attention Module (CBAM) in the neck for enhanced spatial-channel feature representation, and optimizing bounding box regression with a novel Hybrid Intersection over Union (HIoU) loss for small-object detection. To mitigate the scarcity of open E-bike data, we create the Electric Bikes Dataset (EBD), a curated dataset of over 1257 real-road images capturing E-bikes under diverse conditions (low light, occlusion, bad weather, tilted perspectives). The YOLO-DC model is validated on the EBD dataset and in real-world traffic scenarios, achieving a 12.4% improvement in mAP50 (mean Average Precision at IoU = 0.5) and an overall detection accuracy of 98.4% compared to the baseline YOLOv8s model. The YOLO-DC model delivers high real-time performance for E-bike detection and provides a scalable foundation for tracking in subsequent systems, with significant potential to advance intelligent vehicle perception systems and enhance ITS road safety capabilities.