RNN-Based Personalized Activity Recognition in Multi-person Environment Using RFID

Abstract

Personalized activity recognition, which targets specific single-person in multi-person environments, can be very effective in various settings where each person has their own objects and movement patterns. However, most activity recognition researches only deal with general activity recognition, which uses the same model for all single-person individuals. This is because it is difficult to build a customized model for each individual via manual feature engineering. Thus, in this paper, we introduce personalized activity recognition as a new research direction and propose our own approach to build model for each individual using recurrent neural network (RNN). Also, we suggest a graph-based event processing approach to seamlessly collect time-sliced and annotated data. Finally, we construct three kinds of RNN architectures with three different unit types including iRNN, LSTM and GRU, and perform experiments using real dataset. From the experimental results, we conclude that our approach is feasible to build the customized model in real-world for personalized activity recognition.