Occasionally the surroundings can be busy in real-life scan projects. For example, people walking close to the device operator or people opening doors; this could be due to security restrictions on the mapping site or cars passing by during outdoor scans. Many objects can move within the environment that is being mapped, which can often not be prevented. The following section describes the impact of this movement on the device and the data quality attained.
Main impact of moving objects on Mobile Mapping Systems
The best practices when mobile mapping advise us to avoid any moving objects within the environment when mapping. In certain circumstances this might be hard to achieve. The following describes the impact of moving objects and how they can affect data quality.
Simultaneous Localization And Mapping (SLAM) is an algorithm that allows Mobile Mapping Systems to identify their location and orientation within an environment. From the location and orientation, a trajectory can be built to which scan and image data is added, which enables the construction of a final point cloud. To localize itself correctly, the device must be exposed to a significant amount of uniquely structured objects. Based on the surface shapes and corners of these objects, the algorithm can identify its current position compared to its previous position.
Moving objects add a specific difficulty for such systems; from the viewpoint of the device, it is unclear what is moving within the scene - the object or the device? The estimated location of the device will be incorrect and will be added incorrectly to the current dataset.
If a single object moves within a scene, the SLAM algorithm will probably be able to track its own location, especially if there are other objects around which do not move and are uniquely structured. However, moving objects within the following situations will lead to SLAM drifts or breaks, which can be seen as data overlaps, double walls or shortened/extended corridors:
- featureless environments (tunnels, corridors, open fields or streets...),
- scenes with repeating structures (forests, fences, repetitive building outlines...), and
- more moving than fixed structures.
Examples for this can be the operator passing:
- by walking people in a corridor,
- by moving robots along a fenced production line, and
- by bushes and trees on one side and a car driving by on the other.
If SLAM breaks in such situations, this can seldomly be corrected in post-processing. Instead, datasets need to be cut to exclude such difficult scenes and then aligned manually after separately post-processing each part. It is therefore advised to include considerations on moving objects during project planning and inform all stakeholders and contributors on the impact to the scan.
Effects on the SLAM algorithms will be experienced even if the mapping is paused since despite no new data is added to the processed point cloud, the data will be captured and SLAM will run on it to localize the device.
Point cloud noise
NavVis has developed filter algorithms, which eliminate noise from moving objects as long as these objects keep moving continuously. However, a moving object, which stops moving within the scan image, will be present in the data. This might be a person passing by and then standing still to watch the operator map, a self-closing door being opened and then kept open with the foot or hand, or a car slowing down for a pedestrian.
These semi-moving objects will leave traces in the processed point cloud, most often as noise hovering in the air in otherwise open spaces. In NavVis quality maps, people moving often resemble cloudy white areas within the blue coloring without specific outlines.
If necessary such noise can be cleaned in third party tools from the point cloud when using Desktop Processing.
Moving objects within a scene will always be captured in the panoramas triggered at that point in time. Motion blur can likely be seen on such objects within the images, even if the operator with the device stands extremely still while capturing the panorama. Additionally, these moving objects will of course hide fixed objects within the scene which might be of interest in the scan.
Images that show unwanted moving objects can be hidden by IVION when necessary.
Point cloud coloring
Once moving objects are captured in images (and therefore panoramas), these will as well be used for coloring the point cloud. As a rule of thumb, in each point of the point cloud ,the closest panorama will be selected and the colors of objects will be transferred onto the correct point. If this closest location includes a moving object, its colors will be projected onto the fixed points behind it. This can most often be seen as people or parts of them being projected onto floors or walls, or fuzzy coloring around the outlines of a door opening.
Such coloring issues cannot be corrected after post-processing.
Moving objects within a scanned scene have multiple influences on the data quality and can even break a scan in difficult environments. Therefore, best practices during project planning and scanning should be applied at all times to avoid later problems:
- opening doors in advance and using door stoppers to permanently keep them open,
- cleaning the environment from movable objects within the planned trajectory (e.g., chairs), and
- avoiding busy times, keeping the number of pedestrians and cars within the scene as low as possible.
And most important: pausing a scan if an unwanted object approaches or when a door needs to be opened! A scan is then still vulnerable in SLAM performance, but point cloud noise and coloring issues will be avoided since during pauses of the device the captured data will not be used for point cloud construction in post-processing.
Written in May 2022