Multi-view detection of two-wheeled vehicles on urban roads: a novel dataset and a tailored detector

Abstract Detecting two-wheeled vehicles (TWVs) on urban roads from various perspectives remains a significant challenge. This difficulty arises from the limited availability of dedicated datasets and specialized detectors, resulting in suboptimal detection performance. To address this, we present a new TWV detection dataset, TWVD-XMUT, along with a dedicated detector GLCNet. The TWVD-XMUT dataset comprises 5849 high-resolution images (ranging from 1280 × 720–4032 × 3024 pixels) and 119 300 precisely annotated instances. These images include two primary viewpoints: eye-level and bird’s-eye views, and cover diverse weather conditions, lighting scenarios, and traffic environments. The GLCNet detector integrates a multi-level weight-sharing fusion module to extract shared features across multiple scales. It also incorporates a local density clustering algorithm based on adaptive density estimation to identify high-density regions in images. Furthermore, a dual-branch detection head, consisting of detection and heatmap branches, has been specifically designed to improve small object localization and detection. Experimental evaluations on the TWVD-XMUT dataset demonstrate that GLCNet achieves an average detection accuracy of 84.1% and a false negative rate of 20.3%, surpassing state-of-the-art methods. The dataset will be publicly released at https://github.com/XMUT-Vsion-Lab/ TWVD-XMUT.