A new research paper published on arXiv introduces a one-stage learning framework designed to map monocular roadside-camera images directly to vehicle states within a ground-fixed coordinate frame. This approach represents a significant shift from traditional pipelines that typically separate vehicle detection in the image plane from subsequent geometric post-processing to determine spatial positioning.
Direct Spatial Estimation
Instead of relying on multi-step geometric calculations, the proposed method utilizes features from a pretrained object detector to jointly estimate three critical variables: the ground-plane position, the physical dimensions of the vehicle, and its yaw angle. By integrating these estimations into a single stage, the framework leverages visual features for both detection and direct spatial orientation, reducing the computational overhead associated with traditional post-processing.
UAV-Based Ground Truth Generation
To solve the challenge of obtaining accurate labels for training, the researchers developed a specialized data-collection pipeline. They utilized synchronized video feeds from both a roadside camera and an unmanned aerial vehicle (UAV). The UAV served as a temporary top-view sensing platform, providing precise trajectories and orientations that were then transformed into the ground-fixed coordinate frame. This alignment allowed for the creation of high-fidelity ground-truth labels, essential for training a model to perceive 3D space from a single 2D perspective.
Experimental Validation at Mcity
The framework was rigorously tested using data collected at the Mcity Test Facility. The results demonstrate that the system can successfully recover vehicle trajectories and orientations using only monocular imagery. This capability suggests a scalable path for infrastructure-based perception at urban intersections, offering an alternative to more complex systems like VILAM, which relies on LiDAR measurements from roadside infrastructure to correct visual SLAM errors.
By eliminating the need for separate geometric stages, this framework simplifies the architecture required for smart city infrastructure to monitor traffic flow and vehicle dynamics in real-time.
No comments yet. Be the first!