Search

At crowded event entrances, some pedestrians start pushing others to gain faster access to the events, resulting in dangerous situations. Pushing identification in video recordings of events is crucial for understanding pushing dynamics, thereby managing entrances safely. This talk presents a deep-learning framework to help researchers automatically identify pushing in videos of crowds. The framework consists of four modules: (1) Optical Flow Estimator that uses a pre-trained optical flow model to estimate the dense displacement fields from input video. (2) Wheel Visualization for generating motion information maps from the displacement fields. (3) EfficientNet-B0 Classifier that aims to identify pushing behavior from the maps. (4) A False Reduction and Annotation module; to reduce the number of false identifications of the classifier, annotate the regions of pushing and output the annotated video. We used five real-world ground truth of pushing behavior videos for the evaluation. Experimental results show that the framework achieves 86% accuracy. The framework is open-source and available at https://github.com/PedestrianDynamics/DL4PuDe.

Crowding at the entrances of large events may lead to critical and life-threatening situations, particularly when people start pushing each other to reach the event faster. A system for automatic and timely identification of pushing behavior would help organizers and security forces to intervene early and mitigate dangerous situations. In this paper, we propose a cloud-based deep learning system for early detection of pushing automatically in the live video stream of crowded event entrances. The proposed system relies mainly on two models: a pre-trained deep optical flow and an adapted version of the EfficientNetV2B0 classifier. The optical flow model extracts the characteristics of the crowd motion in the live video stream, while the classifier analyses the crowd motion and annotates pushing patches in the live stream. A novel dataset is generated based on five real-world experiments and their associated ground truth data to train the adapted EfficientNetV2B0 model. The experimental situations simulated a crowded event entrance, and social psychologists manually created the ground truths for each video experiment. Several experiments on the videos and the generated dataset are carried out to evaluate the accuracy and annotation delay time of the proposed system. Furthermore, the experts manually revised the annotation results of the system. Findings indicate that the system identified pushing behaviors with an accuracy rate of 89% within an acceptable delay time.

Helmholtz AI Conference 2023, Hamburg, Germany, 12 Jun 2023 - 14 Jun 2023, At crowded event entrances, some individuals attempt to push others to move quickly and enter the event faster. Such behavior can increase the density over time, which could not only threaten the comfort of pedestrians but also cause life-threatening situations. To prevent such incidents, event organizers and security personnel need to understand the pushing dynamics in crowds. One effective way to achieve this is by detecting pushing individuals from video recordings of crowds. Recently, some automatic approaches have been developed to help researchers identify pushing behavior in crowd videos. However, these approaches only detect the regions where pushing occurs rather than the pushing individuals, limiting their contribution to understanding pushing dynamics in crowds. To overcome the limitations of previous methods, this work presents a novel Voronoi-based Convolutional Neural Network (CNN) approach for pushing person detection in crowd videos. As depicted in Figure 1, the proposed approach comprises two main phases: feature extraction and labeling. In the first phase, a new Voronoi-based method is developed and utilized to identify the local regions of individuals, employing both the video and the associated trajectory data as inputs. It then uses EfficientNetB0 CNN to extract the deep features of individual behavior from the identified regions. In contrast, the labeling phase utilizes a fully connected layer with a Sigmoid activation function to analyze the extracted deep features and identify the pushing persons. Finally, this phase annotates the pushing persons in the video. Furthermore, this work produces a novel dataset using five real-world experiments with their associated ground truths, which is utilized for training and evaluating the proposed approach. The resulting dataset consists of 11717 local regions, of which 3067 represent pushing samples, and 8650 represent nonpushing samples. The experimental outcomes demonstrate that the proposed approach attained an accuracy of 83% and a f1-score of 80%.

Pushing is a behavior that is often used by some pedestrians, especially in crowded event entrances, to gain faster access to events. Such behavior increases the crowd's density, affecting crowd comfort and safety. Real-time detection of pushing behavior is crucial for event organizers; to react to pushing behavior at an early stage, hence avoiding uncomfortable and unsafe situations in the crowd. Recently, some approaches have been proposed to automatically identify pushing behavior from videos of crowded event entrances. However, these approaches could not detect the behavior in real-time or near real-time. Accordingly, in this research, we present a real-time system for automatically detecting and localizing pushing behavior in video frames. As shown in the figure below, the proposed system includes three main modules: 1) Target frames retrieving and processing, which first aims to select the required frames and then extract the entrance area from them. 2) Spatial motion extraction uses deep optical flow based on GPU to estimate the spatial visual motion with speed and direction information from the entrance area. 3) Pushing detection is based mainly on a supervised CNN-based classifier and the extracted motion information; to identify and annotate pushing behavior. We build a dataset from several real-world experiments (videos with ground truth) to evaluate the proposed system. Experimental results show that our system achieves promising performance in terms of accuracy and computational time.