Data Collection Pipeline for Big Interior Registration and Modelling using 3D Sensor

This paper highlights on the analysis of several data collection pipeline for big interior using Structure sensor. Data of big, large interior, collected by 3D sensors such as Structure sensor is very crucial and important as it could be used to develop 3D as-built model, where the model can be utilized for various purposes like maintenance, management as well as renovation work. However, collecting data of big interiors can be challenging as the outcome of the 3D model can be less accurate due to dimensions of the big interior which surpasses the range of the sensor. Thus, there is a need to have a proper planning when collecting 3D data representing big interior, especially rooms with clutter and occlusion due to furniture and equipment. This project concentrates on developing and testing suitable pipeline in collecting data representing big interior using Structure sensor. Three different methods were proposed, tested and analysed, where the interior is modelled using MeshLab. Results show that Method Two, which is wall by wall approach, is the most suitable among the other pipeline proposed. Thus, this method can be utilized by professionals and experts when using 3D sensors like Structure sensor in collecting big, large interior data.


Introduction
Recently, 3D modelling representing interior of a building has gained attention from professionals in Architectural, Engineering, Construction (AEC) area.Due to its various benefits in facility management, building renovation, monitoring and restoration, many researchers and experts have been progressing to create a flexible, fast and affordable solution in creating this 3D model [1] [2].Latest technology and innovation in building construction and monitoring like Building Information Modelling or BIM has increase the interest and development of 3D model representing existing, in-used building as well [3] [4].
Utilizing suitable hardware like laser scanner and 3D sensor is very helpful in collecting necessary data representing the interior in developing its 3D model.3D sensor like Structure sensor provides portability during data collection, hence preferred for 3D interior modelling [5][6].The data could also be used for 3D as-built, which is necessary in BIM as well as building digital twin [7].
Structure sensor, which is a mobile Structure Light Systems (SLS), is developed by Occipital.It needs to be connected to a device (like an iPad) in order to be operated.It is equipped with infrared sensor and laser-emitting diode in capturing depth data.Table 1 summarizes the main specifications of Structure sensor.Due to its specification, low cost and portability / mobility of scanning, Structure sensor has been used to collect 3D data for various applications, from synthetic, random data [8], to medical areas [9][10], apart from interior, room modelling.Commercial buildings like higher institutions and offices often furnished with large rooms and interiors like laboratories, lecture halls and meeting rooms.Study has shown that modelling large, big interior like these can be less accurate without proper data collection process [11][12].This occurs due to the range and dimensions of the big interior exceeding the sensors range.Thus, there is a need to propose suitable data collection approach in utilising these sensors to collect in-used, big interior, due to the fact that these interiors are not only having issues because of their sizes, but usually encounter problems of clutter and occlusion from the furniture and equipment.This paper highlights on the development of suitable approaches for data collection of big interiors using Structure sensor, where analysis will be performed to find out the best approach.

Methodology
Figure 1 shows the flowchart representing the methodology applied for this work.It starts off with designing the suitable approaches for collecting data of big interior.The interpretation and checking whether an interior fall under the big interior can be seen as defined in Figure 2. In this case, the chosen big interior is Robotics Workshop at the Faculty of Electrical Engineering & Technology, Universiti Malaysia Perlis.Figure 3 shows its measurement and conditions of the interior with full of clutter and occlusions, which clearly shows that it falls under the big interior due to its size.From here, 3 suitable approaches were developed, based on knowledge and information from previous research [13], the dimensions of the interior and specification of Structure sensor, as well as feedbacks from professionals like surveyor when collecting building interior data.Other parameters were also considered like complexity of the interior, condition of the interior (clutter, occlusion) and others like lighting, windows and door locations.
The 3 developed approaches are as listed in Table 2, together with their brief method.To enhance further, block diagrams representing respective approach is designed as in Figure 4. Standard point cloud data processing were used to all data using MeshLab in order to register and model the collected data.From here, measurements were collected and compared with the actual measurements of the interior to determine the accuracy.Percentage of error were calculated using Equation (1) to validate the results: Data collection from one wall to another 3 -Division and Wall (DW) Data collection based from wall of division / section of room  Collect data from the corner (starting point) in clockwise direction and proceed to the next section Data collection finished when all sections have been collected

Results and Discussion
Figure 5 shows the registration and modelling results from all data collection approaches.With respect of point cloud data modelling, can be analysed qualitatively (visual / appearance) as well as quantitatively (accuracy measurement).As can be seen from this figure, qualitatively, the COI approach did not able to capture all the interior data due to the sensor's specification which is smaller than the room's dimension.Hence, only the nearby item, which is within the specification, can be scanned and collected, as can be seen in Figure 5(a).Meanwhile, by referring to Figure 5(b) and (c), W2W and DW approaches able to capture the interior, however facing difficulties in registering and modelling the data due to its big interior as well as issues of clutter and occlusion as mentioned in Section 2.
(a) Modelling of data from COI approach (b) Modelling of data from W2W approach (c) Modelling of data from DW approach

Figure 5. The resulting modelling from all data collection approaches
To analyse further, quantitatively, measurements from the model and the actual readings are taken and compared.However, since the first approach of COI did not able to produce complete data of the chosen interior, the analysis for this approach is not been performed.Figure 6 shows the location of the selected measurements for analysis (Point A to B) (Point C to D), while Figure 7 and 8 shows the measurement of the points with respect to the W2W and DW approach, respectively.Table 3 shows the summary of analysis based from the calculation as in Equation (1).Based on the summary as shown in Table 3, Approach 2 which is Wall-by-Wall approach (W2W) gives the best results with minimal error compared to Approach 3 (Division and Wall -DW) for both measurements.This may be due to the existence of wall in DW approach that sometimes can be imaginary instead of real wall due to the data collection design and method, which may lead to significant number of data losses and difficult to be registered / modelled, compared to W2W approach.Thus, it can be summarized that W2W approach can give the best results in collecting data of big interior from 3D sensor.Nevertheless, more suitable data processing methods can be applied in order to obtain a better registration and modelling results representing the chosen interior.

Conclusion and Future Work
In conclusion, three different approaches in collecting data of big interior has been proposed and tested.From here, Approach 2 (Wall-by-Wall, W2W) approach proves to show good results compared to the other developed methods.It is hope that this approach can be used as a standard in collecting data of big interior using 3D sensors.However, more tests can be performed to ensure its successfulness, specifically using other 3D sensors towards other big, complex interior.

Figure 1 . 2 .
Figure 1.Flowchart of the overall methodology Figure 2. Big interior determination flowchart

Figure 4 .
Figure 4. (a) -(c) The developed and tested approaches in big interior data collection

Table 1 .
Specifications of Structure sensor

Table 2 .
The developed, tested approaches

Table 3 .
Summary of results