The validation of the systems is always essential to determine how to use it in different application fields and determine the effectiveness of the systems based on the different contexts.
As is well known and has been already stated, the complexity of the documentation and modelling of CH, as well as being motivated by physical reasons, is reflected in the uses and in the extreme variety of interdisciplinary acknowledgements and comparisons that the products must satisfy.
Although the relationships between the future uses of point clouds/3D model and the validation strategy may not be straightforward, it is crucial to adopt validation rules involving metric quality and other parameters concerning the usability of datasets and to consider the application in a significant sampling that can cope with the uniqueness that the cultural complexes present.
In other words, the validation concept that verifies the specified requirements was assumed to be sufficient for the intended use (ISO/IEC Guide 99:2007 & JCGM 200 2007).
To control the overall metric quality of the ZEB clouds, their reliability was firstly considered based on their single use. The first statistical parameter that contributes to the evaluation of the overall reliability was the accuracy, so some surfaces or clouds derived from more precise measurement systems were considered as ground truth to evaluate the deviation of the studied clouds by means of root mean square error (RMSE). It is also necessary to consider the precision of the clouds, because the precision is linked to the concept of repeatability of the measurements and is normally described by the standard deviation (St.dev) parameter. Thus, for the ZEB clouds, the intrinsic precision that is linked to the acquisition mode has always been considered.
With reference to the previous paragraph, the SLAM system, after calculating the raw trajectories, uses an iterative ICP-like process of automatic cloud-to-cloud profile registration to generate the 3D cloud, and this process has always been controlled using loop paths. In addition, the correspondence of the detected surfaces during the time acquisition has been carefully optimised in a non-automatic way by segmenting the clouds and supplementary cloud matching and optimisation operations. Moreover, even if the ZEB scanner system is provided for the indoor and outdoor environments, which has been studied in the aforementioned literature (Díaz-Vilariño et al. 2017; Nocerino et al. 2017; Thomson et al. 2013; Zlot et al. 2014), the level of accuracy and detail and the noisiness of the clouds are quite different. Therefore, the validation strategy in this work includes datasets acquired in different outdoor and indoor configurations to account for these differences.
Lastly, since the need to obtain multiscale and multicontent heritage models has been ascertained and shared, the possibility of using the ZEB scanner in multisensor surveying configurations has also been investigated. In the validation strategy for this study, the abilities of integration relating to the ZEB clouds were compared with other cloud surfaces acquired with systems that offer different resolutions and accuracies and provide therefore different scales of surveys.
In these cases, the dense clouds to which the ZEB data was integrated were mainly derived from UAV photogrammetry, and the DSM and other photogrammetric products were generated to document the overall set of sites or clouds derived from TLS or CRP. To obtain the reference clouds (using UAV photogrammetry, TLS, or CRP techniques), the usual criteria and pipelines were adopted, and they are not described in this paper. For the generation of photogrammetric clouds, the orientation of the blocks of images and the control of the results occurred through using GCPs (Ground Control Points) and CPs (Check Points), or in the LiDAR terrestrial applications, cloud recording with integrated cloud-to-cloud alignment techniques and references (again GCPs and CPs). “Test dataset presentation” provides the framework of the test datasets based on the selections related to the validation strategy.
Test dataset presentation
The datasets being validated belong to two projects of metric documentation of two vast cultural complexes belonging to different construction periods that can be ascribed to architectural heritage and the archaeological site.
The first site corresponds to the castle of Valperga (Turin) built in a strategic defensive position on the top of a hill and forming a system of palaces and gardens “at the French mode” between the 17th and 18th centuries (see also Chiabrando et al. 2017a, b). The second site is the San Silvestro fortress located in the homonymous archaeological park that includes a territory rich in mines frequented from the Etruscan age (VIII-I cent. BC) until the XX century. Both sites were subjected to a multisensor survey to obtain multiscale models derived from terrestrial and aerial techniques and formed by the integration of datasets at different resolutions and scales.
The sites can be read in the first column of Table 2, while the second and third columns state whether the recorded environments were indoor or outdoor. The (A), (B), and (D) test datasets regarding the inside of a cylindrical tower, an underground ice house, and a mining cave were evaluated using the stand-alone validation. The ZEB point clouds recording the courtyard in Valperga castle (C) and the dataset covering the inside paths of the Rocca of San Silvestro (E) were validated instead in their integration of the whole 3D models of the sites involving UAV, TLS, and CRP clouds.
Table 2 Framework of the selected dataset subject to validation Tower (A)
The first dataset considered was the cloud acquired along the narrow and restricted spiral staircase that runs inside a cylindrical tower of the castle of Valperga (Fig. 3a). The ZEB1 acquisition was performed starting from the outside and covering the stairway up to the dovecote and back to the entrance on the ground floor. (ZEB1 dataset 19,000,000 pts./10-min time acquisition).
The structural and material degradation of the staircase and the helical vault of marked constructive interest had already been detected by means of a CRP survey and 3D modelling techniques from the structure-from-motion (SfM) algorithms (Chiabrando et al. 2017a, b). This 1-cm accuracy point cloud has been the reference for evaluating the SLAM-based dataset.
Ice cellar (B)
Historical buildings often contain surprises; from an entrance of one of the buildings surveyed at the architectural scale in the Valperga complex, a long and dark corridor starts and goes under two blocks of the castle, gradually descending and leading to the ice cellar. This path and this buried space seemed excellent to challenge the potential of the SLAM-based ZEB system. Also, in this case, the acquisition using the handheld ZEB scanner was performed by completing the roundtrip, starting from the courtyard (Fig. 4b; ZEB1 dataset 13,900,000 pts./6 min).
Courtyard (C)
The integrated image and range-based survey at Valperga castle was planned with the aim of merging the DSM derived from UAV photogrammetry computed by nadir and oblique images (Fig. 5b), with the dense and very accurate models of the TLS technique by the FARO Focus 3D X120 scanner (Fig. 5a). The use of the MMS ZEB scanner provided the opportunity to evaluate the use in such cases of buildings so densely packed with narrow courtyards that are not suitable for photogrammetric surveys of facades. In addition, the use of the scanner could avoid the heaviness and density in the usage of terrestrial LiDAR clouds. This was an opportunity then to evaluate the use of ZEB clouds in a relevant multisensor survey context. (Fig. 2; ZEB1 dataset 8,200,000 pts./4 min).
Mining cave (D)
Having ascertained that the handheld ZEB system is profitable for the modelling of underground environments such as quarries and mines, the scanner was tested in the medieval mine called Buca della Faina (Fig. 6; ZEB-REVO dataset 41,700,000 pts./25 min).
This mine is strange because it can be traversed in many places only on all fours, and the cloud was only collected with the help of speleologists. The slowness of acquisition explains the point density of this cloud, which counts 42 million points compared to 33 million of the cloud acquired in the fortress (E).
Fortified village (E)
Even the multistratified site of San Silvestro, with its safeguarding landscape and archaeological heritage, was subject to UAV and TLS acquisition and modelling. The densely built area needed detailed 3D LiDAR survey and models because it is subject to consolidation and restoration works since it lies on a slope of the hill presenting landslides.
The TLS survey, operated with a FARO Focus 3D X120 scanner, for the study of construction systems of masonries was necessarily heavy and time-consuming and was therefore an ideal site to test the validity of the clouds acquired by the ZEB system.
The loop acquisition was laid along the entire visit path of the fortress, starting from the entrance to the east and following exactly the ancient ascent to the culminating part of the fortified village and rearing from the other side according to the ancient road that embraces the cone-shaped Rocca.
In 2016, a ZEB-REVO dataset was collected along the whole circular pathway around the Rocca (33,900,000 pts./23 min). In 2017, a second dataset (7,200,000 pts./8 min) was conceived as an integration of the area and is shown in Fig. 7 with an outward and return track.
Metric validation in stand-alone solution
A stand-alone use of point clouds derived by this SLAM-based mobile mapping is conceivable due to the intrinsic metric values of the endorsed concept of the so-called 1:1 scale of ranging measurements returning, and this method is helpful in mapping indoor volumes wherever georeferencing issues are not required. Once known and admitted as acceptable, the reliability verified on the reference model for (A) (the designated CRP model for the indoor space of the tower), the confidence level of the ZEB reconstruction in such indoor scenarios such as the ancient ice house (B) and the archaeomining cave (D) can be reasonably circumscribed and validated for a related scale use.
The problems about the establishment of the relationship with another reference surface are those problems that are primarily necessary to face because of the already clarified non-existence of positioning data and the lack of direct radiometric content in the raw data. Some referencing issues emerge in the operative fulfilment of this purpose. A first solution can be the matching of tie points that are detectable targets on both the 3D point cloud with x, y, and z coordinates measured as references. Complications deriving from this discrete method include the difficulty in recognising and choosing the exact point; however, a statistical evaluation by means of RMSE on many matched point distances can grant the accuracy assessment (Farella et al. 2016) of the action. A more effective solution of point cloud alignment is commonly offered by the control of deviation errors on the performance of an ICP-like algorithm, a so-called cloud-to-cloud, between the ZEB surface on the other point clouds (TLS; CRP). This method, more continuous along the whole considered surface, undergoes possible different precision of the surface characterisation in local details and suffers the common noise factor in the ZEB surface. For these reasons, the alignment was evaluated without and after a process of optimisation (segmentation, outliers cleaning, noise filtering).
The cylindrical tower (A)
The CRP reconstruction inside the tower presented in Fig. 3b is considered the ground-truth surface to validate the ZEB mobile mapping, and this surface has been computed with a controlled error propagation of less than 1 cm on GCPs and about 1.5 cm in the CPs (Table 3).
Table 3 RMSE on the control points in the CRP model computed inside the tower of Valperga castle First, the cloud-to-cloud best fitting alignment on the CRP model of the full raw ZEB surface returned a first mean value of distance deviation of 0.025 m and a St.dev of 0.034 m (reported in the first column in Table 5), showing 67% of the points that actually deviated from the reference model of a value error < 0.02 m and the 26% between 0.02 < error < 0.05 m (Fig. 8a).
The most significant discrepancies > 5 cm are recognisable on the treads of the steps and in horizontal surfaces as windows and doors along the rising trajectory, as shown in Fig. 8b. In particular, the intermediate values of deviation, 2–5 cm in green colour, are spread in the spiral gradient of the stair ceiling and in the north side of the upper volume, as shown in Fig. 8a. The figure also shows the arrival of the outward trajectory as the farthest from the starting point of the scanning.
For these reasons, some crucial issues occur when the sequent validation is performed, separating the roundtrip and going into the ZEB cloud about its own precision evaluation, and these issues are strictly related to the precision of the operating principle on which the system is based. By separating the raw roundtrip, which is helped by the time-marked trajectory, twofold clouds are obtained: the first trajectory, outward (O), and the second trajectory, the return (R), take 6 min plus 4 min, respectively (11,300,000 + 7,700,000 points). The relative alignment comparison, as listed in Table 4, retrieves critical values if the validation is performed without the cited optimisation of the surface points, denouncing a moderate problem of the SLAM-based alignment between outward and return (O&R). For this type of path, Fig. 3a shows the coupling of the alignment with the noisy effect of the raw point cloud in outlier errors, affecting higher values.
Table 4 Statistical values of the deviation analysis of the separated O&R paths for the cylindrical tower An effective optimisation operation, as previously cited, on both O&R provided the more suitable values in Table 4, and these values are provided in Fig. 9 and ensure greater reliability of the SLAM-based profile alignment solution implemented in the ZEB1. The reliability can be strengthened in the observation along the whole tower elevation in Fig. 9a and in the zoomed excerpt in b, where the figure shows the reduced deviation between the O&R mainly bordered in the steps and in the roofing intrados. Deviation distance errors between O&R are statistically represented with a mean value of 2 cm and a St.dev of approximately 3 cm. Specifically, 93% of the points appear with deviation error values under 2 cm (Fig. 9a); however, the 99.5% is under 5 cm. This precision proof cannot be admitted in a 1:50/1:100 scale of representation.
Furthermore, a second validation with the reference CRP model is thus consequently proposed: after both the separation values in O&R and before and after optimisation execution values are presented in Table 5. Flanking is presented in the first column, and the initial values relate to the introduced full raw comparison.
Table 5 Statistical values of the deviation analysis of the O&R paths with the reference CRP model for the cylindrical tower of the Valperga castle The Table 5 values of comparison with the reference show the enhancing obtained in the discrepancies by optimisation operation applied on the O&R ZEB surfaces, in which the discrepancies are affected by outliers and noise errors. The values in Table 5 display, however, a not very remarkable improvement in deviation values from the CRP reference model, proving the intrinsic accuracy. This improvement is limited to a mean of 2.2–2.5 cm with a St.dev of about 2.5–3 cm for both the O&R.
In particular, the outward values can be assimilated to those values representing the full raw comparison. Therefore, the return point cloud reconstruction based on the SLAM profile alignment approach actually benefits from the enclosed environment, and its reconstruction, as shown in the last column in Table 5, shows a better accuracy validation with a 2.2-cm mean value from the reference photogrammetric model.
The ancient ice house (B)
According to the same principle of validation, in the (B) dataset, the verification of deviation error values has been compared before and after the optimisation approach, i.e. with the whole point cloud and surrounding, and then restricted to the underground path that leads to the ice house and to the mapping of the whole hypogeum volume.
The comparison between O&R from the raw surface reconstruction that had the courtyard as a starting point (necessary to connect from a common area the whole 3D ZEB reconstruction, as the following presented) returned the subsequent values in Table 6. The deviation distances, as introduced, suffer from noise and outlier errors, and the optimised process leads to a reduction in discrepancies of approximately 50%.
Table 6 Statistical values of the deviation analysis of the separated paths O&R for the ancient ice house of Valperga castle If deviation maps are examined, as shown in Fig. 10, to compare the raw O&R point clouds, and in Fig. 11 for the optimised and localised surfaces, the establishment of range values of deviation errors distances, as reported in Table 7, allow the interpretation of some key aspects.
Table 7 Statistical results segmented in range error values from the deviation analysis of O&R In Fig. 10, while in the arrival to the ice house, in the farthest point from the initialisation, the ZEB SLAM system aligned both the raw O&R profiles with deviation errors referable to the blue colour for most of the hypogeum area, and the comparison of starting and arriving profiles in the courtyard significantly affected the statistical values, as shown in Tables 6 and 7. The optimisation operation, whose comparison errors are represented in Fig. 11 and listed in Table 7, replaced more significant values supporting the precision aptitude of the system in the 3D mapping of such narrow passages, but some critical areas remain in specific details, in particular, the values in green/red colours. These decreases of precision are identifiable, for example, in the ceiling of the starting portion of the passage towards the ice house, at the beginning of the scanning process; in some parts of the volume shown in Fig. 11a, in which these decreases are due to the space occupied by unrelated masses differently responding to the lightwave signal; in the ventilating outlet elements; and in some vertical planes (Fig. 11a) during the descending passage, which are differently modelled by the outward and return profiles possibly due to the topographic variation influence in the SLAM-3D reconstruction and due to the uniformity of some tunnel areas.
It is necessary to considerate, however, the blue-cyan colour profusion that is equivalent to a deviation error between O&R < 2 cm and 2 cm < err<5 cm. For the optimised surfaces, the deviation error corresponds to 81% of the points for the former and 16% for the latter, comprising an entirety of 97% of points that deviate in the global SLAM-based alignment of lower than 5 cm.
The mining cave (D)
To validate the ZEB 3D mapping system in such a peculiar framework as the Buca della Faina sample shown in Fig. 6, the reasoning about the results should consider the local condition under which the acquisition has been performed, which is the very impressive number of points during the almost 26 min (nearly 1.7 million/min) of the considerably limited trajectory length (only ~ 100-m roundtrip, almost 0.1 m/s) executed in a very intricate cave shown in Fig. 12, which featured limited accessibility for irregular topography, impervious spaces, significantly harsh surfaces, and reduced light conditions.
The verification of the O&R deviation error on the ZEB points to validate and support the intrinsic precision of the SLAM-based operating system that underwent restrictive conditions was performed in two phases, before and after the optimisation. In this case, with the massive amount of irregular vegetation, it caused a noise error affecting the entrance zone and from the starting point to the arrival closure on the street level.
The first segmentation in O&R parts consists of 14-min plus 12-min paths (22,170,642 + 19,499,539 points), from whose comparison a problem of loop closure occurred, as in Fig. 13, of almost 40 cm. This loop closure distance result is within 8.3% of the point value, as shown in Table 9, meaning a mean deviation error between 0.20 and 0.60 m. The deviation distribution along the whole path is, as shown in the first column of Table 8, defined into 0.2 m with a St.dev of approximately 0.30 m.
Table 8 Statistical values of the O&R comparison subjected to the optimisation process The high values are most likely conceivable because of the diffuse noise errors and of the weight in the statistical evaluation of the densely vegetated area at the entrance.
Based on these considerations, the validation approach tackled the possibility of properly segmenting the cloud and proceeding to the optimisation steps, dividing the data densification inside the mining cave at the entrance from the rest of the cloud outside. The two steps of cleaning and filtering are presented in the central columns of Table 8 and have been conducted parallelly to both segments, differentiating the parameter values and tailoring them according to the area (in, out). In the last step, the validation considers the deviation results only related to the O&R for the surface into the cave, and the deviation is presented in the final column of Table 8.
Statistical results on deviation errors from the explained comparison have been grouped into simpler range values, in which the results are shown in Table 9, allowing better appreciation of the optimisation result improvement in the ZEB 3D reconstruction. Figure 14 presents the graphic representation of the in solo deviation analysis, whose values are reported in Table 9. The challenging 3D survey into the Buca della Faina reported a final evaluation on its own accuracy and established its confidence level for metric purposes: the obtained value is a mean error on O&R based on the only cave of about 5.5 cm with a St.dev of 7 cm.
Table 9 Statistical values of deviation errors grouped into ranges related to the main optimisation steps Metric validation in the multisensor survey
The challenging aspects related to the use of this kind of spatial data, whose own precision has been investigated and accuracy evaluated in the previous section, are now faced based on its compatibly and integrability with other multisensory data on a multiscale survey organisation. The use of DSM coming from TLS with higher accuracy and UAV photogrammetry with significant 3D completeness and continuity is used in this phase and integrated to the ZEB reconstruction to evaluate their complementarity or exchangeability.
The Valperga castle courtyard (C)
The integration of the Valperga castle surveyed volumes is performed with LiDAR scans and UAV photogrammetric DSM.
The confidence level is based on the reliability of their metric reconstruction declared in Table 10 for the UAV photogrammetric reconstruction and Table 11 for the TLS and corresponded to an admitted scale of representation between 1:50 and 1:100.
Table 10 RMSE on control points in the UAV photogrammetric reconstruction of Valperga castle Table 11 Metric control for the Valperga castle scans: the accuracy validation on the LiDAR point cloud registration shows a mean error of about 1 cm on target check points and a mean value of 4 mm on the clouds comparison Two samples of 12 × 7 m segments have been selected and optimised, as shown in Fig. 15a, b: 450,000 pts. for the ZEB segment and 5,600,000 for the LiDAR DSM, which is more than 12 times denser than the ZEB one. From the comparison analysis shown in Fig. 15c, the detected deviation errors with mean values of almost 2 cm (2.3 cm St.dev), as reported in Table 12, confirm the possibility to control the accuracy of the SLAM-based system in an outdoor scenario between a 1:100 and 1:200 scale of representation, in which 99% of points deviate from LiDAR less than 5 cm.
Table 12 Statistical values divided into errors ranges corresponding to Fig. 15c The problem shown in Fig. 15c indicates the deviation (green) increasing with the height, which is due to a kind of systematic error related to the mechanical system of distribution mode of laser rays in ZEB1, as confirmed in Cadge (2016) (this has been improved in profile uniformity coverture by the ZEB-REVO system). However, the tangible limited descriptive capabilities intrinsic in the ZEB system, as expected, are localised in architectural details and edges, as shown in Fig. 16, and they become rounded and approximated. Analytical range values display errors of ± 1 cm (green) for 72.7%, almost 20.7% for + 1 cm < error < + 2 cm and 6.3% for − 1 < error < − 5 cm.
In the proposed validation, the TLS DSM served as a basis for the configuration and referencing of a system of roundtrip ZEB scans starting from initialisation in the courtyard setting and directed to the several castle volumes in Fig. 17.
All the SLAM-based mappings belong to the LiDAR DSM as reference and deviation error values are presented in Table 13. These error values are related to the raw alignment via the cloud-to-cloud method and initial control of deviation distances. With better control of error propagation in such a kind of articulated 3D mapping via optimisation procedures of the single clouds, as cited, the accuracy related to the 1:200 scale of documentation is a more than achievable and a very reliable goal for the ZEB system.
Table 13 Statistical values of the deviation between the ZEB and the TLS reference in the same courtyard Starting from the integration of 3D data in this multiscale and multisource model, some schematic cut sections, as shown in Fig. 18, prove the possible integration with the aerial DSM by UAV photogrammetry.
These sections can show the added value of this measurement technology to effectively support the spatial analysis of architectural environments based on its verified reliability and scale accuracy and in relation to the more consolidated TLS approach.
The fortified village (E)
The validation performed in the second outdoor architectural complex adds more critical issues to the testing of the ZEB system in 2016 because the wide area on which the trajectory has been conducted (~ 6000 m2), and in the loop mode (~ 660 m), i.e., the circular closed path, avoids the roundtrip intended because the return path is along the same outward path.
Due to the critical issues clarified, a second dataset in 2017 was collected in an ~ 2000-m2 area and in the roundtrip loop mode (~ 450 m) Fig. 7.
The reference surfaces employed in this phase are the UAV photogrammetric DSM, whose planimetric and vertical reliable accuracy are reported in Table 14, and the LiDAR scans are restricted to the rooms at the entrance area of the village, whose metric control is reported in Table 15.
Table 14 RMSE on control points in the UAV photogrammetric reconstruction of the Rocca Table 15 Metric control on LiDAR scans in the Rocca The problem of referencing and validating the ZEB scan laid on the entire area of the village was performed in a circular loop. First, Fig. 7 shows the quality-marked trajectory superimposed to the width of the SLAM-based point cloud. The best working areas are generally those areas that have been travelled twice.
If a deviation analysis is performed on the comprehensive highly detailed UAV DSM, as shown in Fig. 19, to validate the behaviour of the ZEB surface on such a kind of reliable reconstruction, a crucial problem emerged, as reported in the form of statistical analysis in Table 16. It is reasonably attributable to the drift error affecting the circular path mode in such an irregular topographic setting, where open spaces and narrow passages alternate, straining the SLAM-based alignment algorithm operation. In fact, the analysis of the whole fitting verification retrieved higher values distributed in the path. By choosing the correspondences of some recognisable matching points (i.e., corners) as the example of Fig. 19, the residual errors are locally reduced, as shown in Table 16, confirming the strong need of the system to base its operational effectiveness of featuring geometries in the environment, especially if an outdoor scenario is considered.
Table 16 Statistical values for strategies I, absolute and divided in percentage of points per error ranges, and II, with residual deviation error on matching points According to a profitable optimisation, the approach that can be deployed is based on filtering, segmenting, and separating non-improved parts, noising errors, and encumbering elements, such as rich vegetation inside the village.
The slight resulting variation obtained by this intermediate phase confirmed the intrinsic precision problem in the trajectory deviation, and a focused segmentation approach was performed to verify the drift areas and localise where the SLAM-based alignment underwent a linear deviation along the path in those cases where the trajectory quality denounced criticalities.
For example, in the higher area of the church, as identified in Fig. 20, the greater discrepancy is pinpointed with values that locally reach a mean value of almost 0.6 m with a St.dev of 0.7 m, as shown in Table 17. If the locally segmented surface is aligned by means of matching points in Fig. 21, the control on residual errors on distances provides much better values, with an order of magnitude of 10 cm, as shown in Table 17.
Table 17 Church: statistical values for strategies I and II The segmented sample area in the lower village, as shown in Fig. 20, should be measured, and in this case, the 2017 ZEB 3D reconstruction was validated due to the beneficial O&R trajectory mode employed for updating that area. By flanking the UAV DSM surface comparison, the deviation analysis with a LiDAR point cloud was computed with the TLS approach and added as a reference surface for the validation.
The residual values, which are summarised in Table 18, for this sample confirm the improved SLAM-based registration on the second dataset (2017), reasonably deviating more from the highly detailed TLS surface than the UAV model. This increase can be corroborated by the analysis and validation of the ZEB (2017) surface on the climbing path to the church. Table 19 reports the comparison values between ZEB and UAV DSM for the segmented narrow footpath made by steep steps, as shown in Fig. 22a, with lateral surfaces in a parallel analysis, pre-/post-optimisation. The analysis is finalised to evaluate, through their precision, the descriptive capabilities of both these rapid mapping approaches to replace a demanding TLS deployment.
Table 18 Village: statistical values for surface comparison Table 19 Path: statistical values for surface comparison If the segments of the stair samples are considered and both the contributions of the techniques are evaluated and compared, the ZEB reconstruction, as shown in Fig. 22b, potentially satisfies the scale detail of the mapping purposes from the ground, supporting and even perfecting the aerial photogrammetric purposes. Referring to the comparison of points shown in Fig. 22a, 0.00 m < 91.7% pts. < 0.05 m and 0.05 m < 8.3% pts. < 0.10 m (higher values are related, as visible, to the railing modelled by the ZEB and not reached by UAV).
The optimised ZEB surface can also profitably support the surface triangulation, as shown in Fig. 23, returning a remarkable level of detail, i.e., in step edges.