Description
Optical flow estimates a dense 2D velocity field (u, v) from consecutive image frames, classically through brightness-constancy assumptions (Lucas-Kanade, Horn-Schunck) and increasingly via deep networks (FlowNet, RAFT). In crowd modelling it is the camera-domain analogue of crowd-flow-estimation and is the standard ground-truth source for evaluating CSI-derived flow fields on overlapping camera-CSI deployments.
When it's used
- Camera-based crowd-flow ground truth
- Evaluation reference for CSI / radar flow estimation
- Integration into hybrid camera-RF crowd pipelines
Limitations
- Requires line-of-sight cameras; useless on its own indoors
- Aperture problem on textureless areas
- Computational cost of dense flow at high resolution