A novel deep learning architecture and MINIROCKET feature extraction method for human activity recognition using ECG, PPG and inertial sensor dataset

Rohit Kumar Bondugula, Siba K. Udgata, Kaushik Bhargav Sivangi

Abstract: The research in human activity recognition has gained prominence in various applications, including healthcare, medical, and surveillance. The earlier popular techniques which relied on images or video sequences to perform classification are susceptible to noise, line of sight, and light conditions. The wireless wearable sensors provide a robust alternative to these techniques for data collection and classification. Towards this, we propose the application of Minimally Random Convolutional Kernel Transform (MINIROCKET) for feature extraction on sensor data. The extracted features are then used by classifiers for activity recognition. To this end, we employed two publicly available datasets containing heart rate sensors and motion sensor data on various activities. Further, we showed that the application of MINIROCKET requires significantly less computational time compared to other existing models.