[feat] Optimize clustering

Goal

The goal of this MR is to improve the clustering result, such as remove false positive (tree, leaves, wall et etc in the scene).

Change

• Add function ExtractClustersWithRegionGrowing to grow region on a voxel grid map of motion points. A struct Voxel and a struct ClusteringGrid are added.

• Add function InitClusteringGrid to initialize parameters of ClusteringGrid and create the BackgroundMap. A background voxel grid is created during the initialisation phase to check the validity of a cluster. If 50% of a cluster's voxel can be found in the BackgroundMap, this cluster should be considered as background not a moving object.

• The output of clusters' information is now in the order of the clusterID.

Result

Demo

Here is a demo with a velodyne dataset

Improvement for single cluster for one object.

The top image is the previous result that a person is extracted as two clusters. The bottom image is the improved result.

Improvement for false positive

The top image is the previous result that a part of wall (fixed object) is detected as a cluster of moving object. The bottom image is the improved result.

Parameters

Parameters for hap dataset

Background map (voxel grid)

Runtime

livox-hap average runtime
Motion estimation	38.9ms
Create background map	8.8ms
extract cluster with region growing	1.58ms

Future improvement

• Separation of a big cluster which contains more than one object.
• Fix axis-z for bounding box