LSTM-RNNs combined with scene information for human activity recognition
Wenhui Chen, Carlos Andres Betancourt Baca, Chih-Hao Tou · 2017
Developing an accurate activity recognition system with sensor readings is a challenging task due to the fact that sensors are subjected to a variety of errors and the nature of human activities is dynamic and uncertainties. Recent studies have shown that machine learning approaches can effectively classify human activities. In this study, we analyze sensor readings from accelerometers and gyroscopes using long short-term memory (LSTM) recurrent neural networks to identify human activities and present a location-aware approach to improve the recognition accuracy. The location information is obtained from analyzing images captured by a wearable camera based on a pre-trained model of the deep convolutional neural networks. With the location information, some unlikely activities can be ruled out, leading to better recognition accuracy.