Integrating Sparse Learning-Based Feature Detectors into Simultaneous Localization and Mapping-A Benchmark Study.

Area of Science:

• Robotics and Computer Vision
• Autonomous Systems

Background:

• Simultaneous Localization and Mapping (SLAM) is crucial for autonomous navigation.
• Visual SLAM (V-SLAM) uses image features like keypoints and descriptors for pose estimation.
• Deep neural networks (DNNs) offer learned features as an alternative to handcrafted ones.

Purpose of the Study:

• To integrate sparse learned features into a state-of-the-art SLAM framework.
• To benchmark handcrafted versus learning-based feature approaches in V-SLAM.
• To evaluate pose estimation performance and resource usage.

Main Methods:

• Replaced ORB detector and BRIEF descriptor in ORBSLAM3 with Superpoint (a DNN model).
• Superpoint jointly computes keypoints and descriptors.
• Conducted experiments on three diverse, publicly available datasets.

Main Results:

• Evaluated pose estimation accuracy of learned vs. handcrafted features.
• Assessed computational resource consumption for both approaches.
• Compared performance across different application domains.

Conclusions:

• Learned features offer a viable alternative to handcrafted features in V-SLAM.
• Integration of DNN-based features can enhance SLAM performance.
• Further research can optimize learned features for robust autonomous navigation.