Construction sites mix fast change (workers, machines) with slow change (walls going up, scaffolding coming down). T-ConSG is a LiDAR-only temporal scene graph for that setting. Its nodes are free-space regions, static objects, dynamic objects and events, with spatial-relation edges and “event participation” edges. It is built incrementally from geometric evidence: free space and static objects persist, dynamic objects are tracked, and when the accumulated evidence changes, the graph records an explicit installation or removal event. So the graph holds both what is moving now and how the site has changed.
Findings
- Three application studies rather than one benchmark: dynamics-aware navigation in a bricklaying testbed with a Unitree Go1 and a Velodyne LiDAR, construction progress monitoring on ConSLAM (five sequences recorded about once a month on one floor of a live site), and online change detection in a simulated site.
- The graph ran in real time on the robot’s onboard CPU in all experiments.
- On ConSLAM it recovers most of the ground-truth free space across the four sequences used, while static-object recall varies more.
Limits
- The abstract and results give no comparison with another scene graph or change detector, so it is hard to place.
- Geometry only. Object classes come from geometric evidence, not from a semantic model, so “what changed” is “a block of points appeared”, not “a door was fitted”.
Relevance
T-ConSG makes change an explicit part of the graph, as ChronoGraph does for action-caused transitions, and it needs no camera or semantic model, unlike the metric-semantic map of Khronos. It is a useful case for dynamic-scenes, because construction sites force the split between short-term motion and long-term change that most indoor datasets avoid. See long-term-autonomy.
Links
- arXiv
- Subfield: Semantic mapping