Crowd attributes estimation using support vector machine and deep learning with multi-source sensor fusion.

Abstract

Unfortunate tragedies have previously been the result of high-density human crowds or pedestrian flow. In addition to, crowd behavior as a reaction to an incident aggravates the complexity and disruption of human flow, resulting in possible trampling and crushing situations. Therefore, it is important to monitor such crowd motion for danger warning and prevention. In this study, a frame work was established to provide continuous monitoring and estimation of crowd flow and load on pedestrian bridges, with particular focus on high crowd density enhancing operation safety. A main innovation under sensing instrumentation is the employment of structurally mounted Fiber Bragg Gratings (FBG) Fiber Optic Sensors (FOS), in conjunction with individually held wearable sensing devices incorporating Inertial Measurement Unit (IMU). Furthermore, the approach added innovation under machine learning employment, primarily Convolutional Neural Networks (CNN) along with conventional Support Vector Machine (SVM) algorithms thus generating crowd estimation models from gathered sensors’ data. The concept was validated using experimental measurements on two phases based on crowd replication scenarios on a scaled test bridge. Generated machine learning models demonstrated effectiveness in crowd attribute classification for flow activity and load characterization, along with regression model for load estimation. Multi-modal sensor fusion at the input and feature level was further applied on strain and acceleration data collected enriching the machine learning models, thus enhancing system efficiency and robustness against noisy and time shifted input data. The results showed that the monitoring solution to be highly effective with peak testing accuracy for single class flow activity classification at 98percent, multi-class flow and load characterization classification at 91percent, and percentage error for load estimation regression reaching a minimum of 9percent.