Date of Award

5-2025

Document Type

Project

Degree Name

Master of Science in Computer Science

Department

School of Computer Science and Engineering

First Reader/Committee Chair

Dr. Qingquan Sun

Abstract

In healthcare applications such as disease prevention, sleep quality evaluation, and patient monitoring, bed posture recognition is essential. Using pressure sensor arrays placed on top of or embedded in mattresses, this study investigates the application of deep learning models for non-invasive posture classification. Although they have been widely employed, traditional machine learning approaches like support vector machines (SVM) and k-nearest neighbors (KNN) sometimes struggle with feature extraction and real-time performance necessitating considerable processing resources. I implemented a model using conventional approaches to get over these restrictions, then fine-tuned it using the following deep learning architectures for bed posture recognition: ResNet-50, EfficientNet-B0, MobileNetV2, EfficientNet-B4, and ResNet-101.

Our approach involves pressure sensor-based bed posture detection entails identifying and evaluating patterns of pressure distribution through piezo sensor based detectors. When lying on bed in different positions, such as supine, lateral, or fetal, these sensors record changes in the pressure that various body parts exert. A 2D pressure map is created from the gathered data, in order to be understood by machine learning or deep learning models, data is flatted into 1D later utilized to classify the position. Because every position has a different pressure signature, the model can reliably differentiate between them. which capture pressure distribution patterns corresponding to different body postures. The training process includes Feature Extraction and Initial Training, Fine-Tuning for Maximum Accuracy, Evaluation and Performance Optimization.

Among the evaluated models, EfficientNet-B4 and ResNet-101 demonstrated the highest classification accuracy, benefiting from their deep feature extraction capabilities and optimized architecture. MobileNetV2 exhibited the best trade-off between computational efficiency and classification accuracy, making it suitable for real-time applications on embedded systems. Fine-tuning strategies have significantly improved model performance, with accuracy gains of up to 5-10% compared to baseline pre-trained models. The integration of the Nvidia GPU significantly reduced training time while maintaining high classification accuracy increasing the epochs, making it a viable, convenient platform for real-time, on-device bed posture recognition applications.

Share

COinS