Abstract: In the mountainous regions of Xinjiang, traditional manual survey methods for dangerous rock masses are often restricted by the complex and steep terrain. To improve the efficiency and automation of dangerous rock masses surveys, this study proposes a semi-automatic technique using unmanned aerial vehicle (UAV) for high and steep slopes. This methodology integrates close-range photogrammetry with precise terrain-following flight path planning to generate accurate 3D point cloud models of ultra-high steep slopes. Considering the distinctive shapes of dangerous rock masses protruding from the slope surfaces, this research leveraged CloudCompare software's point cloud segmentation tool to perform semantic segmentation of these profiled blocks. Furthermore, a qualitative assessment of dangerous rock masses is achieved through an analysis of their three-dimensional features. This methodology was applied to the ultra-high slope dam site on the left bank of the Yulong Kashi Hydropower Project. In the test area, four dangerous rock masses were identified (all with stability coefficients lower than 0.9, average around 2000 m³ in volume, with height differences ranging from 7-11m), aligning closely with manual field surveys. The research shows that high-precision slope point cloud models, integrated with rock body characteristics, can effectively detect dangerous rock masses, enhance survey efficiency, and mitigate the inaccuracies associated with manual data collection. This approach holds significant practical value for assessing dangerous rock masses on ultra-high steep slopes.