# Research on target localization and adaptive scrubbing of intelligent bathing assistance system

## TL;DR

This paper presents an intelligent bathing assistance system that improves localization accuracy and adapts scrubbing actions using depth data and real-time detection.

## Contribution

The novel approach combines depth correction algorithms and YOLOv5s for skin detection to enhance scrubbing accuracy in intelligent bathing systems.

## Key findings

- The system achieves a localization error with an average of 2.186 mm and a maximum deviation of 2.23 mm in scrubbing trajectories.
- YOLOv5s enables real-time skin detection despite variations in lighting, occlusion, and skin color.
- Depth correction and cubic B-spline interpolation improve scrubbing precision and adaptability.

## Abstract

Bathing is a primary daily activity. Existing bathing systems are limited by their lack of intelligence and adaptability, reliance on caregivers, and the complexity of their control algorithms. Although visual sensors are widely used in intelligent systems, current intelligent bathing systems do not effectively process depth information from these sensors.

The scrubbing task of the intelligent bath assist system can be divided into a pre-contact localization phase and a post-contact adaptive scrubbing phase. YOLOv5s, known for its ease of deployment and high accuracy, is utilized for multi-region skin detection to identify different body parts. The depth correction algorithm is designed to improve the depth accuracy of RGB-D vision sensors. The 3D position and pose of the target point in the RGB camera coordinate system are modeled and then transformed to the robot base coordinate system by hand-eye calibration. The system localization accuracy is measured when the collaborative robot runs into contact with the target. The self-rotating end scrubber head has flexible bristles with an adjustable length of 10 mm. After the end is in contact with the target, the point cloud scrubbing trajectory is optimized using cubic B-spline interpolation. Normal vectors are estimated based on approximate triangular dissected dyadic relations. Segmented interpolation is proposed to achieve real-time planning and to address the potential effects of possible unexpected movements of the target. The position and pose updating strategy of the end scrubber head is established.

YOLOv5s enables real-time detection, tolerating variations in skin color, water vapor, occlusion, light, and scene. The localization error is relatively small, with a maximum value of 2.421 mm, a minimum value of 2.081 mm, and an average of 2.186 mm. Sampling the scrubbing curve every 2 mm along the x-axis and comparing actual to desired trajectories, the y-axis shows a maximum deviation of 2.23 mm, which still allows the scrubbing head to conform to the human skin surface.

The study does not focus on developing complex control algorithms but instead emphasizes improving the accuracy of depth data to enhance localization precision.