A Data-Driven Approach for GPS Trajectory Data Cleaning

In this paper, we study the problem of GPS trajectory data cleaning, aiming to clean the noises in trajectory data. The noises can be generated due to many factors, such as GPS devices failure, sensor error, transmission error and storage error. Existing cleaning algorithms usually focus on certain types of noises and have many limitations in applications. In this paper, we propose a data-driven approach to clean the noises by exploiting historical trajectory point cloud. We extract road information from the historical trajectories and use such information to detect and correct the noises. As compared to map matching techniques, our method does not have many requirements such as high sampling rates, and it is robust to noises and nonuniform distribution in the historical trajectory point cloud. Extensive experiments are conducted on real datasets to demonstrate that the proposed approach can effectively clean the noises while not utilizing any map information.