Application of IndRNN for human activity recognition: the Sussex-Huawei locomotion-transportation challenge

We propose a human activity recognition method using Independently Recurrent Neural Network (IndRNN) on the spectrum for the Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge 2019. The proposed method takes advantage of FFT and IndRNN to obtain short-time and long-time features, respectively. To be specific, since the signal obtained by Smartphone-sensors is strongly periodic, it is first processed into spectrum by FFT in a one-second sliding window to obtain short-time features. Then, the spectrum of 21 overlapped window extracted from 5-second data is processed by IndRNN in order to explore the correlation of FFT spectrums of all windows to obtain longtime features for the final action classification. To obtain a phone-position independent model, the training data of three phone locations (bag, hips, torso) is used to train the IndRNN model, which is further fine-tuned with half of the validation data in hand position (a relatively small amount of data which can be obtained in practical applications). The proposed method achieved 82.6% accuracy of validation in hand position, which have been submitted to SHL recognition challenge as "UESTC_IndRNN".