Point Cloud Filtering with PCL: Extract, Crop, and Clean
When working with point clouds in robotics (e.g., perception pipelines in ROS 2), raw sensor data is almost never used directly. It’s typically too dense, too noisy, and too large. The Point Cloud Library (PCL) provides powerful tools to remove unwanted points, restrict spatial regions, and clean outliers.
1️⃣ Removing a Subset of Points with ExtractIndices
Sometimes you already know which points you want to remove. For example:
-
Points belonging to the robot body
-
Points classified as ground
-
Points flagged by a segmentation step
PCL’s ExtractIndices lets you either keep or remove a specific index set.
1
2
3
4
5
pcl::ExtractIndices<pcl::PointXYZ> extract;
extract.setInputCloud(filtered_cloud);
extract.setIndices(to_remove); // provides the subset.
extract.setNegative(true); // keep everything NOT indexed
extract.filter(*filtered_cloud); // overwrite original cloud
Use this when: you’ve already segmented something. You want precise removal.
2️⃣ Spatial Filtering with CropBox
A CropBox filter keeps only points inside a 3D bounding box.
In your example, the boundaries are defined as:
1
2
3
4
5
6
7
8
9
Eigen::Vector4f tmp_max(cloud_crop_max_depth_,
2.0 * target_standoff_distance_,
cloud_crop_height_ / 2.0,
0.0);
Eigen::Vector4f tmp_min(cloud_crop_min_depth_ + camera_to_base_link_offset_,
-2.0 * target_standoff_distance_,
-cloud_crop_height_ / 2.0,
0.0);
Why 4D vectors?
PCL uses Eigen::Vector4f because internally it operates in homogeneous coordinates. The fourth component (w) is typically 0.0 for points.
3️⃣ Cleaning Noise with StatisticalOutlierRemoval
StatisticalOutlierRemoval (SOR) removes points that are statistically inconsistent with their neighborhood.
-
For each point:
-
Find its k nearest neighbors
-
Compute the mean distance to those neighbors → did_idi
-
-
Compute global statistics:
-
Mean μ of all distances
-
Standard deviation σ
-
-
Remove points satisfying:
