guest ::
| Home > Publications database > Artificial Intelligence-based Early Pushing Detection in Live Video Streams of Crowds |
| Home > Publications database > Artificial Intelligence-based Early Pushing Detection in Live Video Streams of Crowds |
| Conference Presentation (After Call) | FZJ-2023-05241 |
; ;
2023
This record in other databases:
Please use a persistent id in citations: doi:10.34734/FZJ-2023-05241
Abstract: Entrances of crowded events are often set up as bottlenecks for several reasons, such as access control, ticket validation, or security check. In these scenarios, some pedestrians could start pushing others or using gaps among crowds to reduce their waiting time. Such behavior doesn’t only limit the comfort zones but also leads to threatening people’s safety. Early detection of pushing behavior may assist security and organizers in making decisions on time, enhancing the comfort and safety of the entrances. Unfortunately, existing works reported in the literature to detect pushing in crowds are limited and have not satisfied the early detection requirements. For instance, Lügering et al. [1] developed a manual rating system to understand when, where and why pushing appears in video recordings of crowded entrance areas. To overcome the limitations of manual analysis, Alia et al. [2] proposed an automatic deep-learning system for pushing detection. However, this system does not meet the requirements of early detection. To fulfill the early detection requirements, we present an Artificial Intelligence framework for automatically identifying pushing in the live camera stream in real-time. Our framework consists of two main components: The first component uses a pretrained deep optical flow model and color wheel method to extract the movement of pixels from the live stream of a crowd and represent this information visually. The second component includes an adapted and trained EfficientNetV2B0 model, which extracts deep features from the motion information, and then identifies and annotates pushing patches within the live stream. We created a labeled dataset using five real-world experiments [3] with their associated ground truths to train the adapted model and evaluate the framework. The experimental setups mimic the crowded event entrances, and two experts based on the manual rating system [1] created the ground truths. According to the experimental results, our framework identified pushing patches with an accuracy of 87% and within a reasonable delay time. --- References: [1] Üsten, E., Lügering, H. & Sieben, A., Pushing and Non-pushing Forward Motion in Crowds: A Systematic Psychological Observation Method for Rating Individual Behavior in Pedestrian Dynamics, Collective Dynamics, 7 pp. 1-16, 2022. --- [2] Alia, A., Maree, M. & Chraibi, M. A Hybrid Deep Learning and Visualization Framework for Pushing Behavior Detection in Pedestrian Dynamics, Sensors, 22, 4040, 2022. ---[3] Pedestrian Dynamics Data Archive hosted by the Forschungszentrum Jülich, P. Crowds in front of bottlenecks from the perspective of physics and social psychology, 2018.
|
The record appears in these collections: |
PDF