A Crowd Behavior Identification Method Combining the Streakline With the High-Accurate Variational Optical Flow Model

Xiaofei Wang,Jie Zhang,Lei You,Zishu He,Rui Sun,Jun Hu

doi:10.1109/access.2019.2929200

Abstract

A crowd behavior identification method is proposed by combining the streakline based on fluid mechanics with a high-accurate variational optical flow model in this paper. Firstly, calculated by the high-accurate variational optical flow model, the streaklines are used to acquire the crowd motion trajectory information. The angular histogram and the regions of interest in the scenes are obtained by calculating and clustering the dasymetric dot maps of the starting and ending points of the trajectory, and then, combining the dasymetric dot map and angular histogram information to analyze whether there are specific crowd behaviors in the regions of interest, and thus to identify different types of crowd behavior in such scene. Finally, experimental comparison and analysis are made to verify the effectiveness and accuracy of the method proposed in this paper.

Highlights

Crowd behavior analysis has always been an important research topic in the field of computer vision, yet the occlusion between high-density crowds makes it difficult to identify and track individual targets in crowded scene
DASYMETRIC DOT MAPS OF THE STARTING AND ENDING POINTS OF THE TRAJECTORY Using the streakline based on the high-accurate variational optical flow model described in Section 2, we can obtain a large number of crowd motion trajectories in the scene
(2) Judgement of Lane and Arch: for each trajectory of the clustering center, firstly, the orientation angle of a certain number of sampling points is calculated by (13), and the angular histogram is counted according to the calculated orientation angle, in which the interval length is 10 degrees, and there are 36 intervals

Summary

INTRODUCTION

Crowd behavior analysis has always been an important research topic in the field of computer vision, yet the occlusion between high-density crowds makes it difficult to identify and track individual targets in crowded scene. Stationary Crowd Groups [5], [6], Social Network Model [7], Sparse Reconstruction Cost (SRC)[8], [9], Structured Trajectory Learning (STL) [10], Deep Spatiotemporal Perspective [11], Top-Bottom Hierarchical Clustering Strategy [12], have been proposed to analyze crowd motion pattern and they work well in some specific scenes. These methods focus on a certain kind of crowd behavior and fail to pay attention to different types of crowd behavior.

THE STREAKLINE COMBINING THE HIGH ACCURATE VARIATIONAL OPTICAL FLOW MODEL

IMPROVED STREAKLINE

REGIONS OF INTEREST

EXPERIMENTAL RESULTS AND ANALYSIS

ANALYSIS OF CROWD BEHAVIOR IDENTIFICATION

CONCLUSION