TL;DR
This paper introduces GY-SLAM, a novel dense semantic SLAM system designed for dynamic plant factory environments. By integrating a lightweight YOLOv5-based target detection network (GY) with GhostNet, CoordConv, CARAFE, and SE attention, GY-SLAM significantly improves localization accuracy and robustness in the presence of moving objects, while also reducing computational complexity.
What Problem Does This Paper Address?
Simultaneous Localization and Mapping (SLAM) is crucial for robot navigation, but existing systems often struggle in dynamic environments. This is particularly problematic in settings like plant factories, where robots interact with humans and other moving machinery while transporting goods. Traditional SLAM methods can be misled by dynamic objects, leading to trajectory errors or even tracking loss. The paper aims to address this by developing a SLAM system that can effectively detect and eliminate dynamic features in real-time, enabling more reliable navigation and mapping in such complex, ever-changing scenarios.
Methods
The core of GY-SLAM is built upon the ORB-SLAM3 framework, with significant enhancements in two key areas: target detection and dense mapping.
- Lightweight Target Detection (GY Network): The authors developed a custom lightweight target detection network, GY, based on YOLOv5s. This network incorporates GhostNet as its backbone for efficiency. To further boost accuracy and generalization, they integrated:
- CoordConv: To provide explicit positional information to the convolutional layers.
- CARAFE: A content-aware up-sampling operator for better feature aggregation.
- SE Attention Mechanism: To allow the network to focus on more informative feature channels.
- Dynamic Feature Elimination: The GY network's output (bounding boxes and semantic labels of detected objects) is fed into a dedicated "Target Detection Thread." This information is then used by a "Dynamic Feature Elimination Module" within the Tracking Thread to remove ORB features originating from detected dynamic objects, ensuring only static features are used for pose estimation.
- Dense Mapping: To provide a more useful representation for navigation, GY-SLAM constructs pure static dense point clouds. This involves statistical filtering for outlier removal and voxel down-sampling. From this dense point cloud, it generates:
- 3D Octree Maps: For efficient geometric representation and navigation.
- 2D Occupancy Grid Maps: Crucial for collision detection and path planning.
Datasets
- Proprietary Dataset: The authors collected a new dataset specifically for plant factory transport robots. This dataset includes images captured by an Intel RealSense D455 camera and ground truth trajectories from a NOKOV motion capture system. It features common elements in plant factory logistics: humans, robots, and vegetable packages, captured in various dynamic scenarios.
- TUM RGB-D Dataset: This widely-used benchmark dataset was also employed for evaluating the system's performance, particularly on dynamic sequences.
Key Insights and Ideas
- The Synergy of Lightweight DL and SLAM: The paper demonstrates that a carefully designed, lightweight deep learning model (GY) can be effectively integrated into a robust SLAM framework (ORB-SLAM3) to overcome the limitations of dynamic environments without sacrificing real-time performance.
- Beyond Static Assumptions: By explicitly detecting and removing dynamic features, GY-SLAM moves beyond the static environment assumption of many SLAM systems, leading to significantly improved accuracy in real-world, dynamic scenarios.
- Efficiency Through Architectural Choices: The combination of GhostNet, CoordConv, CARAFE, and SE attention in the GY network strikes a compelling balance between detection accuracy and computational efficiency, making it suitable for resource-constrained robotic platforms.
- Dense Maps for Practical Robotics: The generation of pure static dense point clouds, octree maps, and occupancy grid maps provides a more complete and actionable representation of the environment for downstream tasks like path planning and collision avoidance.
Key Results
- Detection Performance: The GY model achieved a high [email protected] of 95.364%, a slight improvement over YOLOv5s. Crucially, it reduced parameters by 43.976%, computational cost by 46.488%, and model size by 41.752% compared to YOLOv5s.
- SLAM Performance (TUM RGB-D Dataset): GY-SLAM significantly outperformed ORB-SLAM3 in dynamic scenarios, showing improvements of up to 92.59% in RMSE for Absolute Trajectory Error (ATE) and over 90% for Relative Pose Error (RPE).
- SLAM Performance (Proprietary Dataset): On their custom dataset, GY-SLAM achieved improvements of up to 28.43% in ATE RMSE and Mean, demonstrating its effectiveness in realistic plant factory conditions.
- Speed Improvements: Compared to YOLOv5s, the GY model achieved a 41.59% improvement in detection speed. The overall GY-SLAM system saw a 17.80% increase in SLAM operation speed.
- Real-time Capability: The tracking thread of GY-SLAM processed frames in under 10 ms per stage, indicating strong real-time performance.
Limitations and Future Work
- Low-Dynamic Scenarios: The paper notes that GY-SLAM can be slightly constrained in low-dynamic environments compared to specialized algorithms like DynaSLAM and DS-SLAM. This is because GY-SLAM eliminates all features in identified dynamic regions, potentially leading to a scarcity of features for tracking in less dynamic scenes.
- Near-Field Dynamic Targets: The authors suggest future work should focus on better processing of near-field dynamic targets, targets that are static but appear to move, and targets that are dynamic but stationary.
- Geometric Information Integration: Integrating more geometric information with deep learning is proposed to improve the system's ability to judge target motion states and distinguish between near and far points.
- Broader Application: The long-term goal is to integrate GY-SLAM into actual plant factory transport robots to support advanced tasks like recognition, transportation, and route planning, contributing to automation in the logistics industry.
Summary updated