Towards Real-Time Human Activity Recognition: A Machine Learning Perspective
Saikrishna Rajanidi, Rajagopal Saminathan, Sebin Beebi Philip, Thangavel Palanisamy, Senthil Kumar Thangavel, Tirumala Kumar Kandukuru · 2024
Human Activity Recognition (HAR) is a crucial component in the development of intelligent systems, ranging from healthcare monitoring to smart environments. While existing HAR systems achieve promising results, limitations arise in handling a diverse range of activities and differentiating each activity accurately. This research addresses these challenges by proposing models to enhance activity recognition accuracy. The paper presents a comprehensive exploration of HAR using a diverse set of machine learning and deep learning models, aimed at improving accuracy and real-world applicability. This approach involves using the HAR dataset from the UCI ML repository which is a popular benchmark dataset for the same. The pipeline includes data preprocessing, handling outliers, and adding noise to better approximate real-world scenarios. To tackle the inherently complex and dynamic nature of human behavior, a wide variety of techniques from Machine Learning models like SVM with K-Means clustering to sequential models in Deep Learning were implemented. Impressive accuracy scores of 93-94 percent were achieved using linear models like multinomial logistic regression and SVM with K-Means clustering. The sequential Deep Learning models like LSTM and GRU produced a decent accuracy of 80-85 percent, but lesser than their Machine Learning counterparts primarily because of limited data samples. The best-performing model in the study was found to be the multinomial logistic regression model. All the models have been extensively evaluated using different metrics and techniques. The analysis revealed some powerful insights regarding the data of each activity and how to further improve the classification, for the sake of better activity recognition.