Application of hybrid gate recurrent unit for in-store trajectory prediction based on indoor location system

In field of location prediction, trajectory recognition is one of the most widely research issues. Since trajectory includes various information such as position, time, and speed, many scientific methods are applied to extracting meaningful features, and discovering valuable knowledges. This paper pays more attention on case study of in-store trajectory, and proposes a series of recurrent neural network (RNN) for location prediction based on trajectory. The trajectory is provided by a indoor location system (ILS) in supermarket, and used radio frequency identification (RFID) technique to collect customer mobility data namely RFID data. After reviewing relatively previous studies, scholars mostly pursue customer segmentation and classification tasks based on trajectory, this paper briefly focus on regression analysis and pattern recognition of original trajectory itself. This paper also includes two improvements of experimental and methodological design. In ILS experiments, we select crossing sections as universal background to filtering customers and their trajectories, and choose fish, vegetable, and meat sections as experiment target, so as to train a general prediction model for heterogeneous shopping behaviors. In methodologies, we propose several RNNs with hybrid gate units on pre-defined trajectory based on RFID data, and also investigate their advances on time-series regression task for trajectory prediction. According to comparative and numerical results, the proposed models show higher performance on in-store trajectory prediction than other benchmark methods and other classic neural networks.