研究目的
To evaluate the effectiveness and accuracy of camera localization using a 3D feature prior map captured by RGB-D and LiDAR sensors in the context of AR and indoor positioning.
研究成果
The proposed camera localization workflow achieves high accuracy (11 cm mean error for original images, 35 cm for independent images) using a prior 3D feature database and deep learning. The alignment of real and virtual cameras facilitates AR application development. Future work will focus on improving robustness with sensor fusion and geometric constraints.
研究不足
The method relies on a prebuilt 3D map, which may not be available in unknown environments. Performance can degrade with significant lighting changes or perspective variations, requiring more feature correspondences and computational resources. The accuracy of 35 cm for independent datasets may not suffice for all high-precision applications.
1：Experimental Design and Method Selection:
The methodology involves building a 3D feature database using RGB-D SLAM with extended bundle adjustment, incorporating depth data for higher accuracy. A camera localization workflow is developed using ORB feature matching and PnP problem solving with RANSAC. Deep learning (PoseNet) is used for initial coarse positioning. An AR registration method aligns real and virtual cameras in a game engine.
2：Sample Selection and Data Sources:
Data is collected from indoor scenes (e.g., a corridor of 57m x 40m) using a Microsoft Kinect V1 camera. Training and testing datasets include consecutive video frames labeled with camera poses from RGB-D SLAM.
3：List of Experimental Equipment and Materials:
Microsoft Kinect V1 camera (resolution 640x480, field of view 42 degrees), Rigel vz1000 terrestrial LiDAR, Velodyne mobile laser scanner, Linux PC with Intel Core i7-6700 CPU, NVIDIA GeForce GTX 1070 GPU, 16GB memory.
4：Experimental Procedures and Operational Workflow:
Capture RGB and depth data with Kinect, build 3D map using RGB-D SLAM, construct feature database with ORB features. For localization, extract ORB features from query images, match with database, estimate camera pose using PnP and RANSAC. Use PoseNet for initial positioning. Align real camera parameters with virtual camera in irrlicht game engine for AR applications.
5：Data Analysis Methods:
Accuracy is evaluated by comparing localization results with ground-truth camera poses from RGB-D SLAM and LiDAR data. Statistical analysis includes mean error calculations and visualization of alignment errors.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
