Research On Crowd Collective Behavior Recognition Method Based On Manifold Density

Collective behavior is a macroscopic behavioral pattern exhibited by continuous and orderly individuals,which is widely present in various population systems such as bacterial colonies and crowd traffic.The identification and analysis of collective behavior is an important research branch in the field of computer vision and computer graphics.It is of great significance for public safety,transportation,architectural design and other application fields.The fundamental task of collective behavior recognition is to extract behavior patterns composed of tracking points with strong behavior consistency from the video sequence according to the behavior characteristics of the extracted tracking points.There are still many difficulties in identifying this behavioral pattern.Collective behaviors include local behavior patterns with behavioral consistency and global behavior patterns with behavioral continuity.The uneven distribution of tracking point density and the shape of behavior clusters bring difficulties to collective behavior recognition.At the same time,the identification has complex interactions.The collective behavior puts forward higher requirements for the analysis of local behavior consistency and global behavior continuity;Further research shows that the existing behavior descriptors and their behavior consistency measures still have limitations for the scenarios with manifold topological structures of group motion behavior.This thesis focuses on the research of recognition and simulation of collective behavior.The main contributions are summarized as follows.(1)A collective clustering algorithm based on manifold density is proposed,which can identify local and global patterns of collective behavior with arbitrary shapes and different densities.Inspired by the manifold topological structure of group motion behavior,a new manifold distance metric is proposed to mine the deep behavior patterns of group motion.Furthermore,the concept of collective clustering density is defined,and the clustering algorithm based on collective density is used to identify groups with local consistent behavior.This strategy is more suitable for identifying clusters with arbitrary shapes.At the same time,considering the complex interaction between subgroups,hierarchical aggregation merging algorithm is introduced to obtain the global collective behavior pattern,which can effectively characterize the global consistency relationship.(2)The existing behavioral feature descriptors and their behavioral consistency metrics still have limitations for measuring the clustering evolution and metrics of group behavior under successive frames.To solve this problem,a learning method based on graph learning is proposed,which measures the clustering of scene groups and analyzes the group recognition.Based on the neighborhood characteristics of cluster motion,a measure of local orderliness of cluster motion is proposed,which is based on the clustering dynamics in the range.In this method,the subject of aggregate motion is defined,and the theoretical method based on graph learning is extended by using path propagation and exponential generating function.Then a learning method for computing scene aggregation degree is proposed.Then a dynamic clustering algorithm is proposed based on the clustering metrics,which can effectively identify groups with different motion patterns and spatial locations in the cluster.The robustness and superiority of the proposed method are further demonstrated by the recognition results and comparative experiments in a variety of colony scenes.