Accelerometer Based Activity Recognition — Biomedical Instrumentation & Clinical Sensing

Abstract — Accelerometer Based Activity Recognition

Accelerometer Based Activity Recognition: Biomedical Engineering, Medical Electronics & Bio-Signal Modeling Integrated Clinical Investigation: Bio-Sensors · Analog Front-Ends (AFE) · Diagnostic Waveforms · DSP Algorithms · ~7945 Words Chapter 1: Electrophysiological Foundations, Action Potentials & Clinical Significance A person’s movement or relative positioning can be effectively captured by different types of sensors and corresponding sensor output can be utilized in various manipulative techniques for the classification of different human activities. This paper introduces a multi-structural architecture with residual networks and a spinal module developed with depth-wise separable CNN blocks, named FusionActNet , for activity recognition using sensor data. The sensor data is initially divided into two primary categories: static (where the human body is immobile) and dynamic (involving movement of the human body). Two residual networks are trained using these superclass data to effectively classify both static and dynamic activities. A spinal module consisting of depth-wise separable CNN blocks, enhanced with a guidance module, is incorporated for selecting the final prediction. Sensor data moves through the collection of these networks, which include two pre-trined ResNets, and a spinal module. After receiving prediction vectors from the static and dynamic models for the specific data, these are combined to create a concatenated vector. In this training process, the guidance module learns to generate a weight that emphasizes any of the two prediction vectors that were previously concatenated based on input data and their characteristics. The proposed model is evaluated using two publicly accessible datasets: UCI HAR and Motion-Sense. Therein, it successfully handled the highly confusing cases of data overlap. Therefore, the proposed approach achieves a state-of-the-art accuracy of 97.35% and 95.35% in the UCI HAR and Motion-Sense datasets respectively.

Sensor Interfacing & Analog Front-End Conditioning

Activity recognition using wearable sensors has been a trending topic of research for its widespread applicability in diverse sectors ranging from health care services to military applications . Modern mobile devices provide abundant sensor data, which are valuable for applications like activity recognition. Various types of sensor data along with image and video data have been employed for recognizing human activity . In this paper, the time series wearable sensor data (e.g. 3-axis accelerometer and gyroscope) are mainly focused on, as they are pretty easy to obtain and can be used to recognize human activity from distant data which are very small in volume and easy to share through internet.

Bio-Signal Waveform Analysis & Simulation Findings

In numerous literature, support vector machine (SVM), and k th nearest neighbor (KNN) classifiers are the most popular ML-based training method for human activity recognition. Furthermore, multiple-stage training through LSTM-CNN-LSTM, ANN , CNN-LSTM , and several deep learning frameworks , are also suggested by some literature.

Scope & Clinical Applications

• Real-time multi-parameter vital signs telemetry (ECG, Heart Rate, SpO2, Temperature, Respiration)
• Low-noise Analog Front-End (AFE) conditioning with active Driven Right Leg (DRL) 50 Hz cancellation
• Digital signal processing in MATLAB: Pan-Tompkins QRS detection, FIR/IIR notch, and Continuous Wavelet Transforms
• Finite element tissue bio-electromagnetics and orthopedic mechanical simulations in COMSOL & ANSYS
• Complete Arduino firmware .INO sketches, circuit schematics, and university dissertation reports