Data filtering is the task of refining data sets by isolating undesired information. In robotics, the data sources are sensors used to examine a robot’s surroundings and/or observe its actions within the operating environment. Typically, autonomous robots are navigating in unpredictable arenas and require precise feedback to coordinate with outside events. Even in the best scenarios, data filtering is necessary. It requires an in depth knowledge of the available and how to overcome their unavoidable shortcomings.
Most undesired sensor data is the product of multiple noise sources. There are a few cases where this isn’t true however. For example, a low-pass filter can be used to isolate the force of gravity on an accelerometer. In either case, when attempting to correct corrupt data it’s important to understand the forces at play.
Noise can be split into two separate categories: measurement noise and process noise. Measurement noise can be described as the observation error generated without any outside influence. Often times this is somewhat predictable and is advertised on sensors as some form of measurement variance. Process noise is error induced by the operation itself. For example, magnetometers are devices used to read the Earth’s magnetic field in order to provide an absolute heading. While this information is useful, it is very vulnerable in magnetically perturbed environments. DC motors used in robotics can cause massive magnetic disturbances under heavy loads. This disturbance will surely cause process noise in the vicinity of a magnetometer.
The most popular data filters in robotics often have multiple functions. Algorithms such as the Kalman Filter and the Complementary Filter are used for sensor fusion as well as refining single data sets. Sensor fusion is the process of combining sensory data from multiple sources to provide accurate representations of a single system. In doing so, levels of uncertainty are reduced that would otherwise be found in systems relying on individual sources. Sensor fusion performance heavily depends on the sensors themselves and how well they augment each other. For example, an inertial measurement unit (IMU) and global positioning system (GPS) are a natural fit since they generally overcome each other’s shortcomings. An IMU will often drift overtime but can provide accurate, short term measurements. GPS data generally follows a Gaussian distribution that’s more useful when observed over longer timesteps. When used as an input to an Extended Kalman Filter, these two sensors can provide a level of reliability that’s viable for most outdoor applications.