Group Detection And Analysis In Crowded Scenes

Group-level crowd behavior analysis is a new and promising method with important applications for the video surveillance and understanding of crowds.However,a specific definition for group in crowd field has rarely been investigated.This thesis proposes a complete and reasonable definition for group in crowd field and presents a fast and automatic group detection method.What' more,we learn the relative importance of individuals in a group by Support Vector Regression(SVR).And then we show that fundamental group-level properties,such as intra-group stability and inter-group conflict,can be systematically quantified by visual descriptors.The main work of this thesis are below.First,automatic and fast density clustering(AFDC)is used to find the group core,which is then refined based on the property of coherent neighbor invariance.This detection method is more suitable to groups with arbitrary shapes and varying densities because the group core is refined with coherent neighbors.Experiments on hundreds of video clips of public scenes showed that the method achieved an excellent detection performance and attractive statistical results.In particular,the number of people in a group exhibits a power-law distribution truncated by an exponential tail;this is significant to understanding crowd scenes and crowd simulation.Then,a relative importance prediction model is formulated,it can not only be able to say which person is more important,but also predict the relative strengths between pairs of people.To assess the importance of a person,we extracted 45 dimensional features.And at last,we build the three-level group model consists of group core,coherent members and the important members.At last,we systematically study the fundamental and universal group properties and design a rich set of group-property visual descriptors,including geometric structure,topological structure,and collective degree,namely collectiveness,stability,uniformity,and conflict.This is made possible by learning a Group Prior by modeling the tracklets with Markov chain.These descriptors convey richer group-level information in comparison to the conventional group size and velocity information.Importantly,these descriptors are scene invariant and robust to public scenes with variety of crowdedness.