CrowdOcc Targets Occlusion in Crowded Indoor Scenes for Quadrupeds

A quadruped-focused vision model scored 46.23 Human IoU while estimating human occupancy in 3D from a monocular view. The result, reported in the CrowdOcc arXiv preprint, addresses crowded indoor settings where people can occlude static scene geometry and predictions of human occupancy can be incomplete or spatially misplaced.

The researchers built CrowdOcc with 25.1K RGB-D frames from 11 indoor scenes. The dataset includes semantic occupancy annotations constructed by separating static and dynamic elements, a process the paper calls static dynamic decoupling.

Semantic scene completion is intended to represent scene geometry and occupancy, including areas affected by occlusion. The paper focuses on the difficulty of producing complete, correctly positioned human occupancy predictions and recovering static geometry when people block parts of an indoor scene.

CrowdOcc combines two components. Normal Guided Scene Geometry Fusion supplements depth-aware lifting with surface-normal cues to improve geometry under occlusion. Human-Centric Sparse Interaction selectively models human-human relationships and local relationships between people and the surrounding scene in 3D.

On CrowdOcc’s scene-disjoint test set, the method reached an overall IoU of 15.80, mean IoU of 11.40 and Human IoU of 46.23, according to the preprint. The researchers describe the performance as state of the art on CrowdOcc and say the test results show generalization to unseen indoor scenes. The supplied report does not include comparative scores from other methods, so the size of any improvement cannot be assessed from it.

The supplied report describes a dataset and perception framework, not a demonstrated crowd-navigation system. It provides no robot walking speed, inference latency, onboard computing requirement, collision rate or navigation success rate, and it does not describe a quadruped autonomously navigating with the resulting predictions. Whether the method can support safe navigation on a moving robot is therefore not established by the report.