Human activity recognition using deep transfer learning of cross position sensor based on vertical distribution of data

Sensor-based human activity recognition and health monitoring are attaining great interest in the eye of the researcher as it maintained the privacy of an individual. A model based on the transfer learning for the vertical distribution of cross position sensor data is proposed in this paper. The whole human body participates during an activity like walking, jumping, running, etc. When someone walk, run, jump, upstairs downstairs his hands and lags both act as per the activity. Existing methods of human activity recognition using sensor data learn from one dataset and transfer that learning for another one but in the proposed work combine learning of accelerometer, magnetometer and gyroscope placed at ankle of the body is used as starting point for the lower arm of the body. combination of sensor data from three sensors (accelerometer, gyroscope, magnetometer) obtain better result as compare to an individual and other combinations. Two publically available datasets are used, i.e. mHEALTH and PAMAP2. The proposed model has been compared to the state of art approaches. The experimental results show that the proposed approach performs batter compare to state-of-the-art methods. Model score 98.48 and 98.63 % test accuracy for balanced and unbalanced mHEALTH dataset and achieved 92.00 and 94.19 % test accuracy for balanced and unbalanced PAMAP2 dataset.