A smartphone-based human activities recognition using novel multi-stream movelets on fusion of accelerometer and gyroscope data and classification using different distance metrics

• Published in: Multimedia tools and applications Vol. 84; no. 25; pp. 30259 - 30280
• Main Authors: Tokas, Pratibha, Semwal, Vijay Bhaskar, Jain, Sweta
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.07.2025; Springer Nature B.V
• Subjects: Accelerometers; Accuracy; Artificial neural networks; Business metrics; Chebyshev approximation; Classification; Comparative analysis; Computer Communication Networks; Computer Science; Data collection; Data Structures and Information Theory; Datasets; Deep learning; Dictionaries; Gait; Gyroscopes; Human activity recognition; Machine learning; Multimedia Information Systems; Neural networks; Parkinson's disease; Performance evaluation; Performance measurement; Sensors; Skewness; Smartphones; Special Purpose and Application-Based Systems; Track 2: Medical Applications of Multimedia; Wearable technology; Accelerometer; Gyroscope; Movelet; CNN; LSTM; Human activity recognition
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-024-20352-2

In recent times, smartphone-based wearable devices are becoming popular for monitoring human activities. Physical activity patterns can be used to identify health problems. Built-in smartphone sensors, namely accelerometers and gyroscopes, can be utilised to collect data for activity recognition. In this research, the fusion of accelerometer and gyroscope data is used to classify lower extremity activities, namely walking, standing, climbing up the stairs, climbing down the stairs, sitting, cycling, running, lying, jogging and two transitional activities: stand-to-sit and sit-to-stand using Movelet method. The Movelet method creates a personalised dictionary for each subject from only 5 seconds of training data. It classifies activity by calculating the distance between the dictionary and test data. In this research work, we have explored different distance metrics such as euclidean, cosine, manhattan, and chebyshev distance. We have considered Accuracy and F1-Score as performance metrics. Moreover, we have also evaluated the performance of the movelet method on different amounts of training data like 3,4 and 5 seconds. We have identified the optimal size of training data for different groups of activities. We have also compared the accuracies obtained by the Movelet method with traditional deep learning models, CNN and LSTM. Our findings show that the Movelet outperforms the CNN and LSTM model by enhancing the average accuracy by 40% and 50.5%, respectively, when trained with the same training data.