Detecting objects in COLMAP point clouds reconstructed from smartphone video faces two obstacles: the reconstructed frame is not aligned with gravity, which distorts heightbased classification, and a single DBSCAN radius either merges unrelated dense structures or fragments sparse regions. This paper combines RANSAC ground-plane alignment, a densityadaptive DBSCAN (DA-DBSCAN) with a ceiling on the pertier radius, and a size-capped merge step that rejoins fragments without re-creating oversized clusters; PCA descriptors serve as a secondary classification cue. As a proof of concept on one smartphone scene (132,831 points) with one reference subject, fixed-radius DBSCAN placed 97.8% of clustered points in three oversized clusters (largest bounding box 326 cubic metres in the aligned frame). The full pipeline reduces the largest box to 23.0 cubic metres (93%) and recovers the subject with precision 0.99 and recall 0.70 (F1 0.82; recall 0.42 without merging). Its partition is identical under rigid rotation (ten seeded alignments, adjusted Rand index 1.000) and it runs in about 2 s on a CPU-only notebook. With default settings, HDBSCAN isolates the subject in only 3 of 10 orientations of the same cloud and otherwise absorbs it into a cluster 18–21 times larger, and OPTICS assigns it to noise at about 80 times the runtime. The results are descriptive: one scene, one subject, no significance tests. The pipeline is unsupervised and training-free.

COLMAP;DBSCAN;Density-Adaptive Clustering;Ground-Plane Alignment;point cloud data;object detection