This paper proposes an adaptive region of interest (ROI)-aware downsampling method for efficient transmission of 3D point cloud data. Existing approaches such as voxel grid sampling (VGS) and farthest point sampling (FPS) reduce data size but often remove important visual information. To address this issue, our method detects a 3D ROI using multi-axis projection and applies different downsampling ratios to ROI and non-ROI areas. Experiments show that the proposed method preserves the visual quality of salient regions more effectively than VGS and FPS, while also providing stable input for object detection models. In CNN-based detection, it achieved an average detection score of 0.594, improving by up to 0.17 over baselines. In Transformer-based detection, it reached 0.871, exceeding existing methods by up to 0.37. These results demonstrate that our method improves the quality and reliability of point cloud transmission under bandwidth constraints, supporting adaptive streaming and contributing to semantic multimedia communications.