Scene-Independent Group Profiling in Crowd

Jing Shao1, Chen Change Loy2, and Xiaogang Wang1

1Department of Electronic Engineering, 2Department of Informaiton Engineering, The Chinese University of Hong Kong.

 

[PDF] [Presentation(ppt)][Presentation(oral video)] [Demo Video] [Video Spotlight] [Dataset & Code]

 

1. Introduction

As shown in Figure 1, groups are the primary entities that make up a crowd, and group dynamics have been widely studied in socio-psychology and biology as the primary processes that influence crowd behaviors.

Figure 1. Groups widely exist in nature: fish school, ant swarm, bird flock, and bacteria colony. Group dynamics contain both intra- and inter- aspect: e.g. bacterial colonies were found to exhibit COLLECTIVE behavior to achieve a common goal, i.e. spreading of diseases; CONFLICT often occurs during comeptition of resources or goal incompatibility, either in fish schools or ant swarm.

This work aims at proposing a universal and fundamental set of group properties and quantifying them with scene-independent visual descriptors, shown in Figure 2.

Figure 2. Crowd behavior can be better understood through inherent intra- and inter- group properties.

The contribution of this work:

Figure 3. Framework.

2. Group Detection

Precise group detection in crowd is challenging due to complex interaction among pedestrians. We argue that pedestrian movements in a scene are intimatedly governed by a finite number of Collective Transition priors. The key idea of group detection is to search for pedestrian groupings that fit well to the discovered priors within a video clip.

Figure 4. (a) Coherent filtering fails to distinguish two subtle groups. We address this problem through discovering (b) a representative anchor tracklet (marked in red) and subsequently (c) a set of seeding tracklets to infer a group-specific collective transition prior. (d) With refinement based on the collective transition prior, two groups are separated.

 

3. Group Descriptors and Its Applications

We formulate a set of visual descriptors to quantify four group properties. The first three descriptors quantify the spatio-temporal evolvement of intra-group structure, whilst the fourth characterizes inter-group interaction. Two applications are used to demonstrate the effectiveness of our descriptors: group state analysis and crowd video classification. Both applications are scene-independent. Details are shown in the demo video (If you cannot access Youtube, please click here):

 

 

4. Reference

If you use our dataset or code, please cite our paper.

Jing Shao, Chen Change Loy, and Xiaogang Wang. "Scene-independent group profiling in crowd." Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR, oral paper), 2014.  

Jing Shao, Chen Change Loy, and Xiaogang Wang. "Learning Scene-Independent Group Descriptors for Crowd Understanding." IEEE Transaction on Circuits and Systems for Video Technology (TCSVT), 2016.  

 

Back to top

Last update: June 18, 2014