Abstract
There is a growing amount of sensor data is available and these sensors are generating enormous amount of trajectory data. A trajectory traces the movement of any object in a sequence of timestamps. Several research studies have examined trajectory data to determine traffic patterns and distributions. However, the quality of this data varies due to differences in equipment, and there is often noise that needs to be cleaned up. With vehicle trajectories where data is continuously captured, a key problem is determining individual trips by differentiating when a vehicle is stopped for a minute or two versus when a trip has ended. Second problem is detecting and fixing (or removing) noisy data. We propose statistical methods to detect and remove outliers and investigate smoothing and segmentation techniques. Specifically, we introduce methods using the Z-Score over speed and angle measurements to clean up the data. We present results on marine trajectory data which differs from work on automobile trajectories because automobiles generally must travel on roads. These cleaned trajectories can be used to understand movement patterns via trajectory clustering and can be visualized to observe different routes. Understanding the noise patterns can also be useful in improving the accuracy of future sensors.