Wrist-worn Physical Activity Recognition: A Fusion Learning Approach

• Published in: 2020 - 5th International Conference on Information Technology (InCIT) pp. 116 - 121
• Main Authors: Kongsil, Katika, Suksawatchon, Jakkarin, Suksawatchon, Ureerat
• Format: Conference Proceeding
• Language: English
• Published: IEEE 21.10.2020
• Subjects: Accelerometers; Activity recognition; Cross-subjects approach; Data models; Feature extraction; Legged locomotion; Machine learning; Pattern recognition; Sensor phenomena and characterization; Wearable sensors
• DOI: 10.1109/InCIT50588.2020.9310980

With the advantages of technology in sensors on wrist-worn devices, many researchers have investigated and utilized the signals to recognize the physical activity in a living context. However, these studies are still faced with challenging problems, especially improving the recognition rate. Without using the recorded sensor data of target subjects, we aim to create a cross-subject activity recognition framework composed of three important models by using the fusion approach of machine learning. Performance was tested on four public datasets with different window sizes. Our experimental results demonstrate that our proposed framework achieves the highest recognition rate when compared with state-of-the-art models. In some datasets, our framework achieves more than 80% success in the recognition rate. In the other datasets, our proposed model can achieve up to 90%. In summary, our proposed framework is capable of being an activity recognition tool for classifying and identifying the physical activities in real world scenarios.