Trajectory fingerprint: one-shot human trajectory identification using Siamese network

Abstract

Extracting identifiable information from human trajectories is a fundamental task in many location-based services (LBS), such as personalized POI recommendation system, irregular human movement detection and privacy protection. Existing studies assume that we have collected sufficient trajectory data for each user, and therefore a classifier could be trained to distinguish the users. However, in many real-world scenarios, due to unregistered users or less active users, human trajectory data is very fragmentary and we can hardly collect sufficient training samples for each user to train the classifier. Moreover, we could hardly define a clear user set for user identification because the set of users are dynamic and changing everyday (there are always new users and inactive users everyday in the real-world human trajectory dataset). Bearing these in mind, we propose an one-shot learning framework for human trajectory identification to handle the insufficient samples and dynamic user set problems. Sliced encoder neural network is designed to encoder the spatiotemporal characteristics and Siamese network is applied to extract the discriminative features for identification. Experiments are conducted on real-world human trajectory dataset to show the advantageous performance of our algorithm.