Various movement patterns have been discovered in trajectories, which are valuable in studying the contextual behavior of tracked objects [15]. Most of the previous studies have been concentrated on detecting of specific patterns in trajectories, such as flock, leadership and convergence [9] or sequential patterns [14]. In practice, movement patterns can be much more diverse and complex, eg, clockwise, and zigzag, which calls for a unified, effective and robust approach for classification.

In deep learning field, convolutional neural network (CNN) have achieved superior performance in classification of image [8], video [6], text [7] and voice data [4]. In terms of vector data, computer vision approaches have been designed to process point set data for shape matching and classification [1, 11, 12, 13]. However, applying CNN for classification of trajectory data is relatively unexplored, which confronts two challenges. Firstly, trajectory data includes two additional pieces of information, namely the connectivity between points and direction, which can be primary features in movement pattern classification. Secondly, both the number of trajectories in a set and points in a trajectory can be variable. They should be considered together with the variations in point positions. To address the above two challenges, this paper proposes a deep learning approach for classifying regional dominant movement pattern (RDMP), which is defined as the movement pattern followed by the majority of a trajectory set within a target region. To achieve this objective, the proposed approach defines a directional flow image (DFI) by mapping a trajectory set to a grid space according to its extent and storing local directional flow in multiple channels at each grid, ie, the pixel of DFI. The benefit is that a trajectory set with the aforementioned variations can be transformed into an image in fixed shape. Subsequently, a CNN called TR-Net is designed for classification of DFI, which is trained on synthetic trajectory data and a considerably high accuracy is achieved. In summary, the approach adds a bridge between deep learning and trajectory pattern classification.