TL;DR
This paper introduces BY-SLAM, a novel visual SLAM system designed to overcome the limitations of traditional systems in dynamic environments. By integrating a new feature descriptor (BEBLID), an efficient object detection network (YOLOv8s-FasterNet), and a refined dynamic feature point elimination strategy, BY-SLAM achieves significantly improved localization accuracy and robustness, particularly in scenes with moving objects.
What Problem Does This Paper Address?
Simultaneous Localization and Mapping (SLAM) is crucial for robots and autonomous systems to navigate and build maps of their surroundings. However, traditional visual SLAM systems often assume static environments. In real-world scenarios, dynamic objects (like people, vehicles, or furniture being moved) can cause significant errors in pose estimation and map generation, leading to tracking failures. Existing methods struggle with accurately identifying and removing these dynamic elements without discarding too many static features, which can also degrade performance. This paper aims to address these challenges by developing a SLAM system that can effectively handle dynamic environments.
Methods
BY-SLAM builds upon the robust ORB-SLAM3 framework but introduces several key innovations:
- BEBLID Descriptor: The standard rBRIEF descriptor is replaced with BEBLID (Boosted Efficient Binary Local Image Descriptor). BEBLID is a deep-learning-based binary descriptor that offers improved matching accuracy and efficiency compared to rBRIEF, which is crucial for maintaining performance when dynamic features are removed.
- YOLOv8s-FasterNet for Object Detection: To efficiently detect potential dynamic objects, the authors integrate FasterNet, a lightweight network, as the backbone for YOLOv8s. This combination speeds up object detection, making it suitable for real-time SLAM.
- DBSCAN for Semantic Mask Refinement: Instead of relying solely on object detection bounding boxes, which can be overly large and include static background, BY-SLAM uses DBSCAN clustering on depth information to generate more precise semantic masks of dynamic objects.
- Epipolar Constraints for Dynamic Feature Elimination: The system combines the refined semantic masks with epipolar constraints to accurately identify and eliminate dynamic feature points. This ensures that only static features are used for pose estimation.
- Dense 3D Map Construction: A new thread is added for dense 3D map construction. This thread filters out dynamic pixels and employs techniques like ghosting elimination and downsampling to create a clean, static 3D map.
Datasets
The authors evaluated their system using:
- TUM RGB-D Dataset: A standard benchmark for SLAM research, containing indoor dynamic sequences.
- HPatches Dataset: Used for image matching experiments, featuring changes in view angle and lighting.
- Self-made Real-World Scene Dataset: Collected using a Microsoft Kinect V1 depth camera to test the system in practical, dynamic scenarios.
Key Insights and Ideas
- The importance of precise dynamic object segmentation: Simply using bounding boxes from object detectors is insufficient. Refining these with depth-based clustering (DBSCAN) is key to avoiding the removal of static features.
- BEBLID's advantage in dynamic scenarios: Replacing rBRIEF with BEBLID not only improves matching but also helps compensate for the reduced number of feature points after dynamic object removal.
- FasterNet's role in real-time performance: Integrating FasterNet into YOLOv8s is critical for achieving real-time object detection, a bottleneck in many dynamic SLAM systems.
- Dense maps without ghosting: The proposed method successfully constructs dense 3D maps that are free from the ghosting artifacts typically seen in SLAM systems that don't handle dynamic objects.
Key Results
BY-SLAM demonstrates significant improvements over ORB-SLAM3:
- Localization Accuracy: On the TUM RGB-D dataset, BY-SLAM achieves an average localization accuracy improvement of 95.53% compared to ORB-SLAM3. In highly dynamic sequences, improvements in RMSE and SD for Absolute Trajectory Error (ATE) and Relative Pose Error (RPE) exceed 80%.
- Robustness: The system shows enhanced robustness in dynamic environments, maintaining good performance even when faced with significant motion.
- Efficiency: The YOLOv8s-FasterNet backbone allows for real-time object detection on the CPU, achieving approximately 29 FPS. The overall tracking time per frame is also competitive, ranging from 48.6 to 52.3 ms.
- Map Quality: Dense 3D maps generated by BY-SLAM are significantly cleaner and more readable, with fewer ghosting artifacts compared to ORB-SLAM3.
Limitations and Future Work
The authors acknowledge that BY-SLAM, being a feature-point-based SLAM system, may struggle in environments with very few features, such as plain white walls. To address this, they suggest future work could involve incorporating multi-modal feature fusion (e.g., point, line, and surface features) to improve performance in textureless scenes and enhance overall robustness.
Summary updated