Privacy-preserved visual SLAM based on a dual-component approach

Abstract The edge-assisted visual simultaneous localization and mapping (SLAM) receives user images, and then estimates and returns 3D point clouds and camera poses. However, this process may lead to the leakage of user privacy, limiting the application of visual SLAM. To address this issue, we proposed a dual-component approach for visual SLAM that preserves privacy. Firstly, the privacy-sensitive area of input images underwent masking and inpainting using an image inpainting method. The method is based on object detection with multi-scale feature fusion and a two-stage image inpainting algorithm. Secondly, the output 3D point cloud was encrypted into a 3D line cloud, where each line underwent random orientation and length normalization. Experiments were conducted on the Oxford Robotcar and KITTI datasets to evaluate the privacy-preserving performance of the dual-component approach. We also assessed its impact on the pose estimation of visual SLAM. The results demonstrate that the approach effectively preserves privacy while ensuring the pose accuracy of visual SLAM.