CNN for human activity recognition on small datasets of acceleration and gyro sensors using transfer learning

Abstract

This paper describes an activity recognition method for the Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge by Team TDU-DSML. The CNN model reported in our 2018 SHL Challenge was adopted. 5-second FFT spectrogram images from all axes of acceleration and gyro sensor data were treated as input data. We confirmed that a multiple-sensor input model combining acceleration and gyro sensors improves the recognition rate. However, there was insufficient training data in the SHL dataset for the target sensor labeled as Hand. To overcome this difficulty, a transfer learning method was applied to the pre-training model from other sensors labeled as Hips and Torso. After evaluation from all combination of sensors, the transfer learning model from Acc_norm and Gyr_x for Hips and Torso had the best recognition rate of 82.1% at Hand in the submission phase.