Advancing Fall Detection in an Autonomous Bus - Examination of LSTM Technique
Amey Ajit Dakare, Yanbin Wu, Toru Kumagai, Takahiro Miura, Naohisa Hashimoto · 2024
With the advent of autonomous public transportation, ensuring passenger safety, particularly in the context of fall detection, has become increasingly important. This study introduces a fall detection system for autonomous buses, transitioning from a Multi-Level Perceptron (MLP) to a Long Short-Term Memory (LSTM) network. The LSTM model addresses the limitations of the MLP in handling sequential data and temporal aspects of human motion, crucial for accurately detecting falls. Our approach involves a two-stage process: initially employing a 2D multi-person pose estimation network, integrated with the SORT algorithm for effective tracking, and then processing the pose data through an LSTM network. The dataset, compiled from an actual bus environment, includes various fall and non-fall scenarios to ensure robust model testing. The LSTM network demonstrates improved performance over MLP model, particularly in handling complex movements.