Convolutional and LSTM Neural Networks for Sensor-Based Activity Classification and Novel Class Detection
이지유 · Open Access System for Information Sharing (Pohang University of Science and Technology) · 2018
Activity recognition aims to recognize human activities, building a model to label segments in multivariate time series collected from multiple sensors. Most of methods address a closed set recognition problem where the prediction is made within pre-specified labels which the model is trained on. For real-world applications, however, a novel class which has not seen in the training phase may be provided to the model. In this thesis, we present a method for activity recognition which integrates a novel class detection into convolutional and LSTM classifiers, enabling automatic detection of novel classes as well as classification of pre-specified classes. To this end, we combine the activity recognition with the LSTM-based temporal anomaly detection by adopting two convolutional-LSTM neural networks which cooperate together, where one is trained as a classifier and the other is trained to produce class-specific prediction errors that are used to assess the likelihood of novel class. Experiments on the Skoda and the OPPORTUNITY datasets demonstrate the high performance of our deep learning method for novel class detection in activity recognition.