研究目的
To correct for bias in lidar-only motion estimation using a learned approach based on Gaussian process regression, improving odometry accuracy with minimal computational overhead.
研究成果
The learned bias correction using Gaussian processes significantly improves lidar odometry accuracy, reducing errors by approximately 10% across multiple datasets with minimal computational overhead. This approach demonstrates the potential of machine learning to enhance classical estimators and can be extended to other motion estimation pipelines, such as visual odometry.
研究不足
The technique relies on hand-picked input features, which may not generalize well across different datasets without retraining. It is evaluated primarily on urban environments and may not perform optimally in other settings. Computational cost, though low, could be further optimized, and the method assumes biases are uniformly accumulated over time.
1：Experimental Design and Method Selection:
The study uses a Gaussian process (GP) regression model to predict biases in motion estimation. Inputs are high-level features from point-cloud geometry, and outputs are predicted biases applied as corrections to classical state estimators. The methodology involves training the GP model offline with ground truth data and applying corrections online.
2：Sample Selection and Data Sources:
Datasets include the KITTI odometry benchmark and lidar data collected around the University of Toronto campus, totaling over 50 km. Ground truth is provided by systems like Applanix POS-LV.
3：List of Experimental Equipment and Materials:
Velodyne HDL-64E lidar, Applanix POS-LV system for ground truth, MATLAB Statistics and Machine Learning Toolbox for GP regression, libpointmatcher and libnabo libraries for point-cloud processing.
4：Experimental Procedures and Operational Workflow:
Point-clouds are downsampled to keypoints based on intensity and planar features. Scan matching is performed using iterative methods like ICP variants. GP model is trained on features derived from geometry (e.g., distribution of normals), and predictions are used to correct odometry estimates in real-time.
5：Data Analysis Methods:
Odometry errors are evaluated using relative pose changes over windows of frames, compared to ground truth. Performance is measured by percentage errors over path segments (100m to 800m), with statistical analysis via cross-validation and hyperparameter optimization in MATLAB.
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
