Abstract: With the extensive application of unmanned technology in open-pit mines, the integrity of LiDAR point cloud data is crucial for the safe operation of mining trucks. A systematic study was conducted on typical point cloud defect patterns and their repair methods in response to the point cloud defect problem caused by the complex environment in the unloading area. The threat posed by the point cloud cavities in the open-pit mine unloading area to the retaining wall recognition and road surface perception was effectively resolved. First, a geometric occlusion model for the defect of LiDAR point cloud is established, and the correlation between the obstacle occlusion probability, radar parameters and spatial density is quantified to reveal its physical mechanism. Based on the actually collected point cloud data of unloading areas, six typical types of cavities are statistically selected, and their defect patterns are analyzed in detail combined with multi-dimensional features: Type I (road surface slope change), Type II (scattered rock and soil accumulation), Type III (retaining wall gap), Type IV (retaining wall structural defect), Type V (rock and soil collapse), and Type VI (dust occlusion). The first four types originate from terrain occlusion, while the latter two are caused by dynamic high-density occluders, which fills the gap in the existing research on specific scenarios in mining areas. A composite repair strategy based on terrain features is proposed for different defect characteristics. The road surface and retaining wall point clouds are segmented through dynamic grid projection and principal component analysis. The block plane fitting repair (BPFR) method is adopted to achieve efficient repair of the road surface. The three-dimensional structure of the retaining wall is accurately restored based on the Poisson reconstruction method and the concave hull constraint. Finally, the repair results are merged. Experimental results show that while ensuring repair efficiency, the composite method outperforms comparative methods in terms of normal vector consistency (NC=0.99815), root mean square error (RMSE=0.04971 m), and global deviation (SSE=75.92), and can adapt to complex terrains and dynamic occlusion scenarios in unloading areas. This study verifies the effectiveness of the proposed method in real mining area data, improves point cloud integrity and environmental perception accuracy, provides reliable data support for downstream tasks of unmanned mining trucks, and has significant engineering value for promoting unmanned operations in open-pit mines.