Local and higher-level pedestrian flow extraction from overhead monocular video based on following and side-by-side relations
This study proposes a method for extracting pedestrian-flow connection and separation structures from overhead monocular video in walking spaces where multiple flows are close to one another. Because pedestrian flows are not defined by inter-pedestrian distance alone, distance-based integration can erroneously merge crossing or opposing flows, whereas merging and branching require a representation of time-varying same-flow relations. Foot occlusion in overhead monocular video further destabilises ground-plane position estimates used for inter-pedestrian relation evaluation. To address these issues, this study proposes a hierarchical pedestrian-flow extraction method that first stabilises ground-plane position and velocity sequences by adjusting observation covariance according to the foot occlusion ratio, then extracts local flows from following and side-by-side relations, and finally constructs higher-level flows from front-rear connectivity between local flows. Controlled comparisons with distance-only and distance-and-direction clustering showed that distance-only clustering frequently grouped pedestrians from different nearby flows. Distance-and-direction clustering achieved the highest unweighted mean F1-score of 0.96, whereas the proposed method achieved the highest mean precision of 0.98 and higher precision in merging and branching. Component-wise ablation showed that the components play complementary roles in maintaining same-flow connectivity and suppressing erroneous merging between different flows. These results show that the proposed hierarchy maintains same-flow connectivity while suppressing erroneous grouping of different flows. It also captures flow connection and separation through local and higher-level flows.
Authors
- Ryusei Shigemoto (ORCID: https://orcid.org/0000-0002-4360-554X)
- Ryosuke Tasaki (ORCID: https://orcid.org/0000-0002-3619-4498)
- Hiroki Kimura (ORCID: https://orcid.org/0009-0008-9738-8721)
Institutions
- Aoyama Gakuin University (JP)
Publication Details
- Journal
- SICE Journal of Control Measurement and System Integration
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1080/18824889.2026.2744545
- Primary Topic
- Video Surveillance and Tracking Methods
- Type
- article
- Field-Weighted Citation Impact
- 0.00