Inf3DNet:Multimodal Fusion 3D Target Detection for Infrastructure

The rapid development of Cooperative Vehicle Infrastructure Systems(CVIS) now allows intelligent roadside units to enhance perception in complex traffic scenarios. Challenges include spatial alignment issues during installation, detection flaws, and difficulties in perceiving distant targets. Obstructions by vehicles affecting detection accuracy of traffic participants such as pedestrians and cyclists are also prevalent. To address these challenges, we propose an early fusion 3D object detection network(Inf3DNet). Inf3DNet integrates a Feature Path Extraction and Aggregation Network(FPEA) for enhancing image feature integrity and Dual-Modality early fusion Network(DMEF) for improving semantic expression of mid to long-range target features. Additionally, a Single-stage Multi-scale Detection network(SMD) facilitates efficient detection across various scales. Comparisons with the large-scale DAIR-V2X roadside multimodal dataset demonstrate its effectiveness, with our approach achieving higher detection accuracy in complex traffic scenarios compared to state-of-the-art methods. This sets the groundwork for future developments aimed at reducing computational costs while adapting to the real-time detection needs of roadside 3D targets.