Drone-based inspection has become an effective tool for improving safety and efficiency in construction applications; however, designing flight paths that balance coverage, inspection resolution, and limited flight time remains challenging. Conventional path planning approaches typically apply uniform flight patterns and fixed image resolutions across entire construction sites, leading to redundant scanning in low-complexity areas and insufficient inspection of critical regions. This paper presents an adaptive drone path planning framework for construction applications that integrates fractal image processing with deep learning-based hazard detection.

The proposed approach first captures a preliminary image of the construction site and applies a fractal quadtree algorithm to partition the site into regions of varying spatial resolution based on visual complexity. These partitions are clustered into multiple altitude levels, enabling resolution-aware path planning in which drones are deployed at different heights to efficiently inspect regions with distinct complexity requirements. High-complexity areas are assigned finer resolutions and lower flight altitudes, while low-complexity areas are inspected at coarser resolutions from higher altitudes.

To enable automated safety inspection, a YOLO-based deep learning model is employed to identify construction hazards from images captured by drone-mounted cameras. The detection model is trained offline using labeled construction site imagery and is capable of recognizing multiple hazard types under varying environmental conditions. Simulation results using real construction site images demonstrate that the proposed method significantly reduces the number of required scan locations compared to traditional random walk and zigzag flight patterns while maintaining sufficient image quality for reliable hazard detection. The proposed framework provides an efficient and scalable solution for adaptive drone-based construction site inspection.