guest ::
| Home > Publications database > Vision AI for Crowd Analysis |
| Poster (After Call) | FZJ-2026-00432 |
2025
This record in other databases:
Please use a persistent id in citations: doi:10.34734/FZJ-2026-00432
Abstract: AI-based video analysis has significantly improved the ability to understand crowd dynamics, which is crucialfor effective crowd management and improving public safety. Crowd video analysis includes key tasks such asdensity estimation, counting, pedestrian trajectory extraction, and behavior detection. This abstract presentsthree Vision AI-driven projects initiated and developed at IAS-7: two focused on pushing behavior detection, andone aimed at supporting broader crowd analysis tasks.The first project introduces a deep learning approach for localizing individuals involved in pushing behavior inrecorded crowd videos. Its goal is to support understanding the causes and risks of such behavior. Experiments demonstrate that the approach achieved an accuracy of 85%. As this approach does not meet real-timerequirements, the second project addresses this gap by presenting a cloud-based AI system that detects pushingbehavior in live video streams. This enables timely alerts and supports early intervention by organizers andsecurity personnel to prevent escalation. The real-time system achieved an accuracy of 86%.The third project introduces a novel benchmark designed to support the development of efficient AI applications forvarious crowd analysis tasks in dense, high-risk environments such as railway platforms and event entrances. Itfeatures a diverse dataset of 109,913 manually annotated, tailored explicitly for head detection in crowded scenes.The benchmark includes empirical evaluations of eight advanced deep-learning models. Results show that YouOnly Look Once version 9 (YOLOv9) and Real-Time Detection Transformer (RT-DETR) achieved mean averageprecisions of 90.7% and 90.8%, respectively.
|
The record appears in these collections: |
PDF