GNV2-SLAM: vision SLAM system for cowshed inspection robots

Area of Science:

• Robotics and Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Simultaneous Localization and Mapping (SLAM) is crucial for autonomous robot navigation.
• Traditional SLAM systems struggle in dynamic environments due to issues like feature loss.
• Cowshed inspection requires robust and accurate autonomous navigation systems.

Purpose of the Study:

• To develop an innovative SLAM system, GNV2-SLAM, for dynamic environments, specifically for cowshed inspection.
• To improve the accuracy, robustness, and real-time performance of SLAM systems.
• To integrate a lightweight yet accurate object detection network into the SLAM framework.

Main Methods:

• Proposed GNV2-SLAM system based on ORB-SLAM2, incorporating a lightweight GNV2 object detection network (YOLOv8-based) with GhostNetv2 backbone, CBAM attention, and SCDown downsampling.
• Implemented point and line feature extraction techniques to handle dynamic targets and blurred images.
• Evaluated performance on the TUM dataset and in a real-world cowshed environment.

Main Results:

• GNV2 network achieved 95.19% mAP@0.5, with a 41.95% reduction in parameters, 36.71% decrease in computational cost, and 40.44% model size reduction.
• GNV2-SLAM demonstrated significant improvements over ORB-SLAM2 in dynamic environments, with a 96.13% reduction in Absolute Trajectory Error (ATE) RMSE.
• Achieved real-time performance with single-frame processing under 30 ms and showed superior trajectory consistency in cowshed trials.

Conclusions:

• GNV2-SLAM offers a robust and accurate solution for autonomous navigation in dynamic environments, outperforming traditional methods.
• The system's efficiency in object detection and feature extraction enhances its applicability for tasks like automated cowshed inspection.
• GNV2-SLAM provides a competitive technical solution for advancing the automation of agricultural inspection tasks.