With the advent of digital transformation in the construction industry, drones have become indispensable tools for site supervision due to their agility, diverse payload capabilities, and ability to provide elevated perspectives. In our recent practice on a large-scale urban development project, we systematically applied drone technology to enhance construction site control. This paper details our workflow, focusing on overcoming common challenges such as route planning in complex environments and model distortion caused by significant height variations. Through a series of structured experiments and data analyses, we demonstrate how drone regulation compliance and advanced flight strategies can significantly improve the effectiveness of site management.
Our work is grounded in a real project situated in a densely built urban area, near a major stadium and a metro line. The site covers approximately 69,515 m² with a total floor area of 528,572 m², including five towers up to 127 m in height and a four-story basement. Such a complex environment demands meticulous drone regulation to ensure both flight safety and data quality. We employed a DJI Mavic 3T enterprise drone equipped with RTK (Real-Time Kinematic) positioning and a loudspeaker attachment for real-time communication.
Technical Workflow Overview
The drone-based construction site control follows a structured pipeline: mission planning, pre-flight risk assessment, data acquisition, post-processing, and model application. We first divide the site into distinct zones based on construction progress and obstacles. Two primary zones were identified: Zone 1 was in the initial earthwork phase with relatively uniform elevation, while Zone 2 was in the superstructure phase with a high-rise building already exceeding 100 m. The latter required careful consideration of tower cranes, high-voltage power lines (approximately 70 m in height), and radio-restricted airspace near the adjacent stadium. All these factors fall under the umbrella of drone regulation; we strictly adhered to local airspace restrictions and company safety protocols.
The following table summarizes the key characteristics of the two zones:
| Parameter | Zone 1 | Zone 2 |
|---|---|---|
| Primary construction phase | Earthwork & site leveling | Superstructure (basement to tower) |
| Maximum obstacle height | ~40 m (machinery) | ~150 m (building + tower cranes) |
| High-voltage power lines | North boundary (~70 m) | North boundary (~70 m) |
| Radio-restricted airspace | None | Adjacent stadium (no-fly zone) |
| Drone regulation constraints | Maintain 30 m clearance from power lines | Avoid stadium airspace; keep distance from cranes |
| Flight altitude (above takeoff point) | 90 m | 180 m (dual elevation strategy) |
| Image overlap | 80% front/side | 80% front/side |
| Number of flight lines | 5 (1 nadir + 4 oblique at 45°) | 6 (3 for low zone + 3 for high zone) |
Drone regulation played a pivotal role in determining the flight boundaries. For Zone 2, we had to ensure that the drone never entered the stadium’s restricted cylindrical airspace. We defined a safe buffer of 50 m laterally and 30 m vertically. Additionally, the presence of mobile tower cranes required real-time monitoring; we configured the obstacle avoidance system to a high sensitivity threshold and limited flight speed to 8 m/s to allow timely response. The flight path was computed using a grid-based waypoint algorithm, where the waypoint spacing was derived from the desired ground sample distance (GSD) and overlap.
Route Planning Optimization
The standard approach for oblique photography sets a constant flight altitude relative to the takeoff point. However, when the site has extreme vertical disparities (e.g., deep excavation pits next to tall structures), this leads to severe resolution variation and model distortion. In our case, the high-rise building in Zone 2 was approximately 120 m taller than the surrounding basement area. If we flew at 180 m above takeoff (to cover the building top), the ground resolution for the basement region would be too coarse, and the oblique images of the building sides would suffer from perspective foreshortening and poor texture. This is a typical problem where drone regulation alone cannot solve; we had to redesign the mission plan.
We divided Zone 2 into two sub-zones: Sub-zone A for the low-rise area (average ground elevation ~40 m, maximum structure height 60 m) and Sub-zone B for the high-rise area (building top at ~150 m). For each sub-zone, we set a different focal plane:
- Sub-zone A (low area): Nadir flight at 180 m (above takeoff), oblique flights at 180 m with focus plane at 40 m elevation.
- Sub-zone B (high-rise): Nadir flight at 180 m, oblique flights at 180 m with focus plane at 150 m elevation (building top).
This dual-focus strategy ensures that the oblique camera is tilted towards the target elevation, minimizing the angular distortion. The theoretical foundation can be expressed by the relationship between the ground sample distance (GSD) and the flight parameters:
$$
GSD = \frac{H}{f} \cdot \frac{p}{s}
$$
where \( H \) is the camera’s height above the target, \( f \) is the focal length, \( p \) is the pixel size, and \( s \) is the sensor width. For oblique photography, the effective \( H \) varies across the image. To maintain a uniform GSD over the building facade, we need to adjust the tilt angle \(\theta\) so that the principal axis of the camera intersects the target elevation. The required tilt angle can be approximated by:
$$
\theta = \arctan\left(\frac{H_{drone} – H_{target}}{D}\right)
$$
where \( H_{drone} \) is the drone’s altitude, \( H_{target} \) is the elevation of the focus plane, and \( D \) is the horizontal distance from the drone to the building. In our mission, we used a fixed tilt of 45° for oblique passes, but by shifting the focus plane we effectively changed the area of optimal sharpness. The resulting improvement in model quality was quantified by comparing the root mean square error (RMSE) of checkpoints on the building roof before and after optimization.
The following table presents the distortion metrics for the two approaches:
| Metric | Single-altitude approach | Dual-focus optimized approach |
|---|---|---|
| Number of images acquired | 1,200 | 1,800 |
| RMSE on roof checkpoints (cm) | 6.8 | 2.1 |
| Texture completeness on top surface (%) | 72 | 98 |
| Model distortion artifacts (presence) | Significant (blurred edges, holes) | Minimal |
| Processing time (hours) | 4 | 6.5 |
Drone regulation also influenced the choice of flight timing. We conducted all flights between 10:00 and 14:00 local time to minimize shadows, as per local airspace regulations that limit flight hours to daylight conditions. Additionally, we obtained a temporary airspace waiver from the stadium authority to allow flights within a 200 m radius of the stadium’s edge, subject to a maximum altitude of 180 m and mandatory real-time telemetry sharing.
Modeling and Distortion Correction
Upon data acquisition, we performed photogrammetric processing using Structure-from-Motion (SfM) algorithms. The key steps include feature matching, bundle adjustment, and dense point cloud generation. One common issue is radial and tangential lens distortion, which is corrected using camera calibration parameters. For our drone camera (24 mm equivalent focal length), the distortion parameters were obtained from a pre-flight calibration target. The correction model is:
$$
\begin{aligned}
x_{corrected} &= x (1 + k_1 r^2 + k_2 r^4 + k_3 r^6) + 2p_1 xy + p_2 (r^2 + 2x^2) \\
y_{corrected} &= y (1 + k_1 r^2 + k_2 r^4 + k_3 r^6) + p_1 (r^2 + 2y^2) + 2p_2 xy
\end{aligned}
$$
where \( (x, y) \) are the distorted coordinates, \( r^2 = x^2 + y^2 \), \( k_i \) are radial distortion coefficients, and \( p_i \) are tangential coefficients. After undistortion, we applied a dense matching algorithm to generate a 3D point cloud. However, the dual-altitude dataset introduced inconsistencies in point density between the top and bottom regions. To merge the two sub-zone point clouds seamlessly, we used a weighted interpolation scheme based on inverse distance weighting:
$$
Z_{merged}(X,Y) = \frac{\sum_{i=1}^{N} w_i Z_i}{\sum_{i=1}^{N} w_i}, \quad w_i = \frac{1}{d_i^2}
$$
where \( d_i \) is the Euclidean distance from the query point to the \( i \)-th point, and \( Z_i \) is the elevation. This blending reduced boundary artifacts significantly. The final mesh model was generated with an average resolution of 2 cm on the building facades and 5 cm on the ground.
Application of the 3D Model
The reconstructed model served multiple control purposes:
- Linear and area measurements: Using built-in measurement tools, we extracted distances between column centers, floor heights, and excavation volumes. The accuracy was validated against total station surveys, achieving a mean deviation of 3.2 cm over a 50 m span.
- Progress comparison: By repeating the same flight plan weekly, we generated time-series orthophotos and models. The difference in volume of earthwork between two epochs was computed using:
$$
\Delta V = \sum_{cells} (Z_2 – Z_1) \cdot A_{cell}
$$
where \( Z_1 \) and \( Z_2 \) are the digital elevation models at Times 1 and 2, and \( A_{cell} \) is the cell area. This allowed automatic progress tracking, flagging delays when the achieved volume fell below the planned curve.
- Safety monitoring: The model enabled virtual inspection of high-risk zones, such as steel structure connections and crane operator visibility. We integrated the loudspeaker to issue real-time warnings when unsafe worker behaviors were detected via the live video feed, enhancing drone regulation compliance for on-site personnel.
Conclusion
This practice demonstrates that drone technology, when combined with meticulous drone regulation adherence and optimized flight planning, dramatically improves construction site control. The dual-focus approach solved the model distortion issue arising from extreme height differences, achieving sub-5 cm accuracy on high-rise structures. Tables and formulas quantify the improvements, providing a repeatable framework for similar projects. Future work will integrate AI-based object detection to automate violation spotting, further strengthening the role of drones in construction management.