A novel one-stage approach for pointwise transportation mode identification inspired by point cloud processing

Transportation mode identification is fundamental for transportation planning and management. With the popularization of ubiquitous GPS-enabled devices, leveraging travelers' GPS trajectories to infer transportation modes becomes a cost-effective and appealing approach. The prevailing two-stage framework of transportation mode identification usually suffer from the inevitable segmentation errors in the first stage, and can hardly achieve real-time inference. The existing one-stage framework models either require multi-source data as input or solely enable fixed-size features, which may need to be further improved. In concern of the similar data structure and semantic segmentation task for point clouds and GPS trajectory points, this study proposes a novel one-stage method to directly predict pointwise transportation modes by introducing and improving PointNet, which is a widely used deep learning network in point cloud processing. Specifically, 1D convolution and pointwise pyramid pooling structure are embedded into the original PointNet to capture local features in various granularities for better distinguishing similar transportation modes. Moreover, a post-processing algorithm is further proposed to refine the pointwise classification by taking the nearby consistency into account. Experiments on the GeoLife dataset show that the proposed method achieves an accuracy of 0.849 in identifying five transportation modes, including walk, bike, bus, car, and train. Comparisons reveal that the proposed method significantly outperforms other state-of-the-art methods in terms of local context extraction capability, computational efficiency, and prediction accuracy, making the proposed approach more efficient and effective in practice.