LAUSR

ABSTRACT Detecting movement patterns with complicated spatial or temporal characteristics is a challenge. The past decade has witnessed the success of deep learning in processing image, voice and text data. However, its application in movement pattern detection is not fully exploited. To address the research gap, this paper develops a deep learning approach to detect regional dominant movement patterns (RDMP) in trajectory data. Specifically, a novel feature descriptor called directional flow image (DFI) is firstly proposed to store the local directional movement information in trajectory data. A DFI classification model called TRNet is designed based on convolutional neural network. The model is then trained with a synthetic trajectory dataset covering 11 classes of commonly encountered movement patterns in reality. Finally, a sliding window detector is built to detect RDMP at multiple scales and a clustering-based merging method is proposed to prune the redundant detection results. Training of TRNet on the synthetic dataset achieves considerably high accuracy. Experiments on a real-world taxi trajectory dataset further demonstrate the effectiveness and efficiency of the proposed approach in discovering complex movement patterns in trajectory data.