Crowd Behaviour Analysis

Automatic understanding of human behavior is increasingly becoming an important domain in computer vision, due to its wide applications for human action recognition, crowd behavior analysis and, ultimately, in intelligent surveillance. Despite recent progress in this direction, analyzing motion information is still considered a challenging topic in the computer vision community due to the typical issues such as occlusions, background clutter, low video quality and impredictable camera motion. This is even more challenging in crowd scenarios, where people movements are governed by complex social interactions, group and individual dynamics, size and context of the crowd scenario.

Crowd Behavior Analysis

In this context, we are conducting an extensive research by proposing novel computer vision frameworks to detect abnormal events such as panic and acts of violence from video sequences. Our research covers a wide range of computer vision techniques, by introducing novel physic and heuristic concepts to model motion of pedestrians which are able to uncover a wide range of complex crowd dynamics. For instance, in the picture, it is represented a violence activity in which interaction forces are estimated by using a physics-based approach: red pixels indicate image regions with a high amount of force, while blue pixels indicates regions where low forces occur.

Pedestrian path forecasting is one of the recently emerging applications in visual crowd analysis and modeling. Among attempts put forth so far, only a few ones have considered the undergoing interaction between agents as a key factor in determining their walking trends in a given scene. To this end, we propose an effective framework for pedestrian path prediction in crowded scenes. In the picture, we show the predicted pathways of pedestrians in a crowd environment starting from a model learnned on the previous walking path (red dots belong to the trajectories of pedestrians at training time, while yellow points indicate the predicted paths).

 

Link to the code: [download]

Link to the dataset: [download]
This is a slightly modified version of "The BEHAVE video dataset", with annotations updated by us. For additional information, please contact Sadegh Mohammadi or Pavistech.

 

 

References:

  • S. Mohammadi, A. Perina, H. Kiani, V. Murino
    "Angry crowds: Detecting violent events in videos"
    European Conference on Computer Vision (ECCV), 2016

  • S. Mohammadi, H. Kiani, A. Perina, V. Murino
    "Violence detection in crowded scenes using substantial derivative"
    12th Advanced Video and Signal Based Surveillance (AVSS), 2015

  • S. Mohammadi, H. Kiani, A. Perina, and V. Murino.
    "A comparison of crowd commotion measures from generative models"
    28th Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2015

  • R. Raghavendra, A. Del Bue, M. Cristani, V. Murino
    "Optimizing interaction force for global anomaly detection in crowded scenes"
    13th IEEE International Conference on Computer Vision Workshops (ICCV Workshops), 2011