International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 11 Issue: 05 | May 2024
p-ISSN: 2395-0072
www.irjet.net
ENHANCING HUMAN ACTIVITY RECOGNITION THROUGH MULTIMODAL ENSEMBLE: A FUSION OF TRADITIONAL AND DEEP LEARNING MODELS Eureka D1, Dr. Sridevi Ponmalar2 1Department of Data Science and Business Systems, M.Tech - Data Engineering, SRM Institute of Science and
Technology, Kattankulathur, Chennai - 603 203, Tamil Nadu, India 2Department of Computational Intelligence, Assistant Professor, SRM Institute of Science and Technology,
Kattankulathur, Chennai - 603 203, Tamil Nadu, India ---------------------------------------------------------------------***--------------------------------------------------------------------ranging from healthcare to personalized user Abstract - Human activity recognition (HAR) is a
interfaces[1][12]. Our project addresses the challenge of creating a comprehensive HAR system that surpasses the limitations of existing methods. The 563 features collected in this experiment are derived from the sensor signals (accelerometer and gyroscope) of the smartphone worn by 30 subjects. Time domain characteristics and frequency domain features are the two groups into which they are separated.
challenging task that has many applications in domains such as patient-care, sports, health-care, security, elderlycare, and a variety of other applications. HAR aims to identify the activities of daily living (ADL) performed by a person using sensor data collected from wearable devices or smartphones. Our proposal for this project is an improved HAR system that integrates sensor data with the fusion of Ensembled Traditional ML models and DeepLearning techniques to achieve high accuracy and resilience. We compare each model's individual performance to the ensembled model. We used the HAR dataset from the UCI repository for our model. The data is collected using a waistmounted smartphone with embedded accelerometer and gyroscope sensors to capture the 3-axial linear acceleration and angular velocity of the person(subject). Butterworth low-pass filter is applied to separate the body acceleration and gravity components of the acceleration signal. We use five classical machine learning classifiers: RandomForest, GradientBoosting, ExtraTreeClassifier, K-NearestNeighbors, SupportVectorMachine, as well as DeepLearning models: MultiLayer Perceptron, DNN. We evaluate our system on the publicly available HAR dataset constructed using recordings of thirty participants engaging in six activities of daily living: sitting, standing, lying, walking, walking upwards and downstairs. We compare the performance of our Ensembled system with the individual models' performance and show that our system achieves good accuracy. This system can be used for various HAR applications that require high accuracy and reliability. The next parts provide a brief overview of how our model works, the dataset, and potential enhancements.
Time domain features: These are obtained by applying various statistical functions to the sensor signals in each window of 2.56 seconds. Some of the functions are mean, standard deviation, maximum, minimum, correlation, etc. The time domain features also include the activity label and the subject identifier for each window. Frequency domain features: These are obtained through the sensor signals' use of a Fast Fourier Transform (FFT) and then computing various spectral measures such as energy, entropy, frequency bands, etc. The frequency domain features also include the angle between the gravity vector and each of the sensor axes. The total number of features is 561 (from the time and frequency domains) plus 2 (activity label and subject identifier), which makes 563 features in total.
1.2 Challenges in Existing HAR Systems In the ever-evolving landscape of technological advancements, Human Activity Recognition (HAR) stands as a crucial domain with applications ranging from healthcare monitoring to smart environments[1][12]. However, the existing HAR systems encounter challenges in accurately identifying a diverse array of human activities, often due to limitations in model adaptability and collaboration between traditional and deep learning approaches.
Key Wors: HAR, UCI repository, Ensemble, Traditional ML, Deep Learning, ADL
1.INTRODUCTION 1.1 Importance of HAR in Modern Applications The evolution of wearable devices has opened new frontiers for understanding human behavior. Recognizing daily activities from sensor data is crucial for applications
© 2024, IRJET
|
Impact Factor value: 8.226
|
ISO 9001:2008 Certified Journal
|
Page 170