**4. Results and Discussions**

#### *4.1. Point Cloud Characteristics*

Regarding the evaluation of registration quality, the point cloud characteristics should be thoroughly discussed. Table 3 summarizes the dataset's characteristics in all scans. The findings showed a relatively high average result of RMSE. The first reason behind that is Faro Focus 3D usually has a higher ranging error [+2 mm at 10 m and 25 m each at 90% and 10% reflectivity excluding the noise], according to the data from the manual of Faro Focus3D [47]. The second reason for a higher RMSE is that fewer tie points were not scanned when the scans were conducted initially. As a result, manual point matching was used, leading to a relatively higher registration error, as previously confirmed by [33]. However, the RMSE results in [19] showed a lower RMSE/Scan of 1.68 mm than the registration results conducted in this paper due to the usage of signalized targets on presurveyed site control points.


**Table 3.** Datasets characteristics in the case study.

The noise of the point cloud revealed an average of 3.6 mm, within the threshold of the range noise. The results also showed a minimum overlap of more than 30% for the four scans. The inclinometer mismatch error of all scans also indicated lower results, implying a good-quality scan registration within the sensor specification.

#### *4.2. Point Cloud Assessment*

KNN search algorithms were used to evaluate the quality of the point cloud datasets by extracting specific geometric features (linear index Lλ, planar index Pλ, and scattered index Sλ). Table 4 illustrates the geometric features obtained from the KNN search algorithms (see Section 3.1). The sample set was calculated respectively based on the eigenvalues. The results of a method I revealed that the salience features of the sample points were (linear, planar = 42.9%). However, the salience feature changed to (planar = 50%) and (planar = 74%) in methods II and III, respectively. The feature extraction values in Method III were more accurate than in Method I and Method II due to the use of entropy information that led to less unpredictability of data (more distribution).

**Table 4.** Geometric Features Extraction according to KNN methods.


Therefore, the results showed that the majority of the sample sets were classified as linear and planar with a small index of scattereness, which was reflected in the robust distribution of the datasets. However, previous studies pointed out that the sample set on forests was divergent, where the salience feature changed based on the type of objects. Some objects, such as stems, exhibited a linear index or planar index, while others, such as leaves, exhibited a scatterness index [39,43]. Similarly, in this paper, the results of the geometric features indicated the structure of the sample set where most structural elements were classified as linear or planar.

### *4.3. Three-Dimensional Object Recognition*

The proposed approach's object recognition results were demonstrated using recall and precision rates for two incorporated scans. The recall and precision results achieved exceptionally satisfactory performance, 98%, and 99%, respectively, on average between scans. The minor errors result from objects with only a few points acquired in the scans or temporary objects with a few points wrongly recognized (false negative and false positive rates).

Figure 3 shows the object recognition results obtained from the scan on 25 December 2020, where the foundations were not recognized because the data acquisition date was after the backfilling activities, making the foundations invisible for the laser to recognize. As a result, the foundations were colored red. However, the second-floor columns showed visible progress in the schedule as the work performed exceeded the schedule; hence they

are colored green. Some of these columns failed to be recognized because of occlusions; hence they are colored yellow.
