Detection

The object detection step receives the foreground point cloud as input data. A preprocessing step removes outlier points from the foreground point cloud. Then, a variant of the DBSCAN clustering algorithm is applied to the filtered point cloud to find cohesive segments or clusters.

Parameters

Foreground neighbor filter

See configuration API definition.

Minimum points (default: 3, range: [1 … 30]): The minimum number of neighbors a given point must have within the specified search radius to be classified as an inlier, not an outlier. Combining a low value with a small neighbor radius removes more noise points.
Neighbor radius (default: 0.5, range: [0.1 … 3.0], unit: \(m\)): The radius within which the required number of neighbors must exist for a point to be considered an inlier. Combining a small value with a low number of required neighbors removes more noise points.

DBScan clustering

The clustering step does not have parameters exposed in the WebGUI.

For advanced usage, see the configuration API definition.

Algorithm

Object detection consists of three steps: point cloud filtering, clustering, and post-processing.

Foreground neighbor filter

The radius outlier filter removes outlier points. This specific algorithm considers points as outliers when they do not have a significant number of neighbor points (default: 3) within a configured radius (default: 0.5 \(m\)).

DBScan clustering

DBScan clustering is a density-based spatial clustering algorithm using the Euclidean distance between points. The remaining filtered foreground point cloud is split into cohesive regions with a consistent point density. The point density is defined as a radius (default: 0.2 \(m\)) which is adapted across the point’s distance from the sensor. Additional filtering ensures that detected clusters contain a minimum of points (default: 10).

Limitations

Objects appear merged: When objects in the scene are too close together or overlap (e.g., a person walking in front of a moving car) they may appear as a single merged object. More advanced clustering methods are required to correctly distinguish such objects.
Processing overload: When too many points are contained in the foreground point cloud (e.g., when the sensor is using a static motion detection and is tilted or moved causing all points to appear as foreground), object detection may require a significant amount of time to determine the clusters in the scene.