An Optical Flow- and Machine Learning-Based Fall Recognition Model for Stair Accessing Service Robots

Abstract

One of the reasons for the lack of commercial staircase service robots is the risk and severe impact of them falling down the stairs. Thus, the development of robust fall damage mitigation mechanisms is important for the commercial adoption of staircase robots, which in turn requires a robust fall detection model. A machine-learning-based approach was chosen due to its compatibility with the given scenario and potential for further development, with optical flow chosen as the means of sensing. Due to the costs, complexity, and potential system damage of compiling training datasets physically, simulation was used to generate said dataset, and the approach was verified by evaluating the models produced using data from experiments with a physical setup. This approach, producing fall detection models trained purely with physics-based simulation-generated data, is able to create models that can classify real-life fall data with an average of 79.89% categorical accuracy and detect the occurrence of falls with 99.99% accuracy without any further modifications, making it easy and thus attractive for commercial adoption. A study was also performed to study the effects of moving objects on optical flow fall detection, and it showed that moving objects have minimal to no impact on sparse optical flow in an environment with otherwise sufficient features. An active fall damage mitigation measure is proposed based on the models developed with this method.