Crowd Analysis

Monitoring a human crowd is a tedious task for CCTVs operators especially during social events. In order to help them by providing software dedicated to this task, physicist, computer graphic and computer vision communities formed the crowd analysis research field where they focused their effort on the elaboration of robust methods and algorithms. It is a very broad area that takes interest in studying all kinds of crowds that can be found in real life. Due to the great variety of crowds, one can expect to encounter in real-life scenarios, methods and applications are plenty.

Crowd analysis mainly uses video data of real-world situations, working with features like optical flow for global motion estimation, to perform tasks such as density estimation, anomaly detection or collective behavior characterization among many others.

Our research team focuses on the analysis of very high-density crowds during social events with the following themes :

    • Physics-based dynamic modeling for the detection of emergent dangerous collective motion
    • Optical flow estimation for abnormality detection in a very high-density crowd: By abnormalities, we meant any movement of the crowd that can cause security issues. Our task is to detect that kind of movement during social events such as musical concerts, crowd meeting where people are vulnerable. The crowd analysis mainly based on the optical flow data. Optical flow methods estimate the global and local motion within a scene. The direction and velocity of each pixel are estimated through different technics. By using machine learning algorithms, our aim is to discover the underlying relationship between motions and events. In opposite to sparse crowd, abnormalities detection in a high-density crowd is challenging since it is obviously difficult to obtain precise motion estimation of numerous small, partly independent, self-occluding, non-rigidly moving individuals. We are working on proposing dedicated methods to improve both the optical flow estimation methods for highly dense crowds and abnormal event models.
    • Deep learning-based methods for high-level crowd analysis

Comments are closed.