Background Subtraction

Algorithms seperating the input point cloud into foreground and background.

Table 1. Available fields in blickfeld.percept_pipeline.config.BackgroundSubtraction
Field	Type	Default	Unit
mixture_of_gaussians	MixtureOfGaussians	-	-
mixture_of_gaussians	Use mixture of gaussians for foreground/background detection
static	Static	-	-
static	Use a static tree structure for foreground/background detection.

Mixture Of Gaussians

Dynamic background subtraction using a mixture of gaussians. The algorithm automatically slowly updates the background model if the scene changes. The update rate is controlled via 'exponential_decay'.

Table 2. Available fields in blickfeld.percept_pipeline.config.BackgroundSubtraction.MixtureOfGaussians
Field	Type	Default	Unit
num_initialization_frames	uint32	-	-
num_initialization_frames	The amount of initially received frames which are used to build the background. Reasonable default: 10
exponential_decay	float	-	-
exponential_decay	Controls how fast objects switch between foreground and background. Exponential decay factor. Reasonable default: 0.005
min_weight_threshold_for_background	float	-	-
min_weight_threshold_for_background	Controls how much noise the background/foreground is expected to have. Reasonable default: 0.25
object_awareness	ObjectAwareness	-	-
object_awareness	Configure the object awareness of the background subtraction.

Object Awareness

The dynamic model update can depend on the detected objects. If enabled, the model update is freezed for areas around detected objects.

This can be very helpful to avoid that stalled objects, which stay in the scene for a while (e.g. parked cars) or objects which are following similar paths (e.g. persons in a hallway), cause the generation of an invalid background model. The dynamic model update is though still enabled for small objects or noise points (e.g. caused by vegetation).

To still allow the merging of objects into the scene background, simple rules are available which automatically re-initialize the background model for the area around a stall object. Also, objects can be explicitly merged into the background via separate API commands.

Table 3. Available fields in blickfeld.percept_pipeline.config.BackgroundSubtraction.MixtureOfGaussians.ObjectAwareness
Field	Type	Default	Unit
enabled	bool	-	-
enabled	Enable object awareness
auto_merge_object_duration	float	-	s
auto_merge_object_duration	After an object has been tracked for the given duration, the area of the object is reinitialized. If 0, the auto merging is disabled. Reasonable default: 10 minutes
sizes	repeated ObjectSize	-	-
sizes	Only enable the object awareness for the given object sizes. If empty, it is enabled for all sizes. Reasonable default: OBJECT_SIZE_MEDIUM, OBJECT_SIZE_LARGE.
disable_for_in_background_objects	bool	-	-
disable_for_in_background_objects	Disable for objects which are marked with the in_background property. This can be beneficial in scenes where the continuous background model update works better than the one-shot auto-merge re-initialization. Please note that the shadow detection will not work as expected when this is disabled.

Static

Uses the first received frames to build a static background map of the environment. The algorithm requires the scene to only contain the static parts of the scene on the start of the processing. The background model is not updated automatically afterwards. The ResetBackground grpc method allows to update the static map of the environment.

Table 4. Available fields in blickfeld.percept_pipeline.config.BackgroundSubtraction.Static
Field	Type	Default	Unit
num_initialization_frames	uint32	-	-
num_initialization_frames	How many frames of the first received frames are used to build the background Reasonable default: 10
voxel_size	float	-	m
voxel_size	Size of voxels used for voxelization of the background during initialization. Default: 0.05 m
min_distance	float	-	m
min_distance	Minimum spatial distance to background to detect points as foreground Default: 0.15 m