Analyzing Crowd Behavior in Highly Dense Crowd Videos Using 3D ConvNet and Multi-SVM

Abstract

Crowd behavior presents significant challenges due to intricate interactions. This research proposes an approach that combines the power of 3D Convolutional Neural Networks (ConvNet) and Multi-Support Vector Machines (Multi-SVM) to study and analyze crowd behavior in highly dense crowd videos. The proposed approach effectively utilizes the temporal information captured by the 3D ConvNet, which accounts for the spatiotemporal characteristics of crowd movement. By incorporating the third dimension as a temporal stack of images forming a clip, the network can learn and comprehend the dynamics and patterns of crowd behavior over time. In addition, the learned features from the 3D ConvNet are classified and interpreted using Multi-SVM, enabling a comprehensive analysis of crowd behavior. This methodology facilitates the identification and categorization of various crowd dynamics, including merging, diverging, and dense flows. To evaluate the effectiveness of the approach, experiments are conducted on the Crowd-11 dataset, which comprises over 6000 video sequences with an average length of 100 frames per sequence. The dataset defines a total of 11 crowd motion patterns. The experimental results demonstrate promising recognition rates and achieve an accuracy of 89.8%. These findings provide valuable insights into the complex dynamics of crowd behavior, with potential applications in crowd management.