Monocular 3D Object Detection from Roadside Infrastructure

Cooperative vehicle infrastructure system (CVIS) plays a crucial role in achieving fully autonomous driving. However, Conducting research on infrastructure-side monocular 3D object detection is challenging due to the significant discrepancy in calibration parameters of cameras mounted on different infrastructures. This discrepancy can create ambiguity for detection algorithms. To address this issue, our approach focuses on directly regress 8 vertices of 3D bounding box at image level to mitigate the impact of calibration parameters. During the training and inference process, our method do not need any calibration parameter. The 3D pose and position parameters are obtained after post-processing. We proposed a simple post-processing algorithm to calculate 3D parameters from 8 image-level vertices. And since background from the view of infrastructure remains unchanged, we propose using Gaussian Mixture Model (GMM) branch to generate moving-objects-sensitive (MOS) features. This approach enhances the recognition of objects, leading to our method being termed GMMNet. GMMNet achieves a high mean average precision (mAP) on the DAIR-V2X-I dataset, surpassing other start-of-the-art methods by a significant margin. Furthermore, GMMNet exhibits a greater generalization ability.