Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 133

Fig. 2.20 ( a ) Original image ( left camera ). ( b ) Background subtracted image. ( c ) Full point cloud
obtained with the correlation-based stereo vision technique. ( d ) Reduced motion-attributed point
cloud

Fig. 2.21 ( a ) Over-segmentation and cluster velocities. ( b ) Objects with convex hull

2.3.4.2 Over-Segmentation for Motion-Attributed Clusters

To simplify the scene representation, we apply a hierarchical clustering algorithm,
recognising small contiguous regions in the cloud, based on features like spa-
tial proximity or homogeneity of the velocities. This procedure deliberately over-
segments the scene, generating motion-attributed clusters. By incorporating veloc-
ity information for clustering, we expect an improvement in segmentation at these
early stages of the algorithm, without the need for strong models to ensure separa-
tion of neighbouring objects. For clustering, we apply the complete linkage algo-
rithm to describe the distance between two clusters. The resulting hierarchical tree
is partitioned by selecting a clustering threshold and addressing each subtree as an
individual cluster (cf. Fig. 2.21 a). The criterion for selecting the threshold is the