Condition and Position Estimation Using Spatial Ultrasound and Machine Learning

Abstract

This study addresses the significant risk of bathroom-related injuries, often exacerbated by slippery surfaces and delays in accident detection, particularly as individuals frequently bathe unattended. Traditional accident detection methods, including those based on radio waves, optical sensors, and cameras, have inherent limitations. Although previous research has explored beamforming with low-cost ultrasound array sensors for object detection, its considerable computational load necessitates alternative approaches, such as machine learning. This paper presents a novel position and condition estimation method employing a Transformer-based model coupled with ultrasound array sensors tailored for bathroom environments. The proposed model architecture incorporates an encoder and two distinct decoders to generate binary segmentation masks and depth estimations, complemented by a multilayer perceptron classification block for condition estimation. The model was trained on binary masks and depth images acquired from a depth camera. Input signals received at 227 kHz are processed through cross-correlation with 25 kHz transmitted signals, followed by downsampling. The performance evaluation was conducted in two areas: bathtub and general washroom. In the bathtub area, condition estimation achieved $97.5~\pm ~2.0$ % accuracy, whereas position estimation yielded a mean absolute error (MAE) of $11.9~\pm ~2.9$ cm and an intersection over union (IoU) of $0.735~\pm ~0.022$ . Similarly, for the washroom area, condition estimation accuracy was $96.5~\pm ~0.8$ %, with position estimation resulting in an MAE of $14.1~\pm ~2.0$ cm and an IoU of $0.656~\pm ~0.023$ . Notably, In the context of critical incident detection, the recall rates for drowning and falls were exceptionally high, exceeding $99.3~\pm ~1.5$ and $99.5~\pm ~1.0$ %, respectively. These results demonstrate the significant potential and effectiveness of the proposed method for robust accident detection within bathroom environments.