Estimation of crowd flow and load on pedestrian bridges using machine learning with sensor fusion

Abstract Unfortunate tragedies have been the result of crowd behavior in high-density pedestrian flows as a reaction to an incident aggravating the conditions and resulting in trampling and crushing situations. Therefore, it is important to monitor such crowd motion for danger warning and prevention. In this work, we utilize a recently introduced framework for simultaneous crowd and structural monitoring based on a novel combination of sensing technologies that includes the employment of structurally mounted Fiber Bragg Grating (FBG) Fiber Optic Sensors (FOSs) in conjunction with individually held wearable sensing devices incorporating Inertial Measurement Units (IMUs). We propose effective approaches for the estimation of crowd flow and load on pedestrian bridges using advanced machine learning techniques with sensor data extracted from both structural sensors and wearable devices. A key novelty of the approaches presented is the utilization of sensor fusion at both the input and feature level within the machine learning methods used. The machine learning approaches include Convolutional Neural Networks (CNN) and Support Vector Machines (SVM); they are first applied on each of the sensing sources, FOSs and IMUs, individually and then extended for multi-modal data fusion at both the input and feature levels. Data fusion enriches the learning models and enhances the system efficiency and robustness against noise and input data time-shifting errors. The proposed approaches are validated using experimental measurements from a crowd replication scenario run on a small-scaled test bridge. The trained machine learning models demonstrate effectiveness in crowd attribute classification for motion speed and load designation, along with a regression model for load estimation. Results show that the proposed monitoring solution is highly effective with peak testing accuracy for single class motion speed classification at 98%, multi-class motion speed and load characterization classification at 91%, and percentage error for load estimation regression reaching a minimum of 9%.