Simultaneous Localization and Mapping

Data association is the problem of matching sensor observations (e.g., corner features from a camera) to known landmarks in the map. For each new observation, you must decide which landmark (if any) it corresponds to. Incorrect associations (assigning an observation to the wrong landmark) corrupt the map. Modern SLAM uses statistical hypothesis testing or machine learning for robust association, but ambiguity in repetitive environments remains a major challenge.

Loop closure occurs when the robot returns to a previously visited location. Detecting and constraining loop closure causes the accumulated odometry drift to be corrected. When a loop closes, the map uncertainty (covariance) decreases significantly, and past states can be updated retroactively. Loop closure is essential for global map consistency; without it, long missions accumulate unbounded drift.

Different sensors excel in different environments: LiDAR provides accurate metric measurements and works day/night but is expensive; cameras are cheap and rich in information but sensitive to lighting; IMUs provide short-term drift-free motion estimates but accumulate errors; ultrasonic and sonar work in darkness or underwater. Multi-sensor fusion (lidar + camera + IMU) is most robust.

Standard SLAM assumes a static environment. In dynamic environments (moving people, vehicles, clutter), algorithms must separate dynamic objects from static landmarks. Methods include: (1) temporal filtering (ignore fast-moving features), (2) semantic segmentation (recognize and mask moving objects), (3) dynamic object tracking (track and predict motion of dynamic elements separately).