AUTHOR=Du Xinwu , Li Tingting , Jin Xin , Yu Xiufang , Xie Xiaolin , Zhang Chenglin 

TITLE=GNV2-SLAM: vision SLAM system for cowshed inspection robots

JOURNAL=Frontiers in Robotics and AI

VOLUME=Volume 12 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2025.1648309

DOI=10.3389/frobt.2025.1648309

ISSN=2296-9144

ABSTRACT=Simultaneous Localization and Mapping (SLAM) has emerged as one of the foundational technologies enabling mobile robots to achieve autonomous navigation, garnering significant attention in recent years. To address the limitations inherent in traditional SLAM systems when operating within dynamic environments, this paper proposes a new SLAM system named GNV2-SLAM based on ORB-SLAM2, offering an innovative solution for the scenario of cowshed inspection. This innovative system incorporates a lightweight object detection network called GNV2 based on YOLOv8. Additionally, it employs GhostNetv2 as backbone network. The CBAM attention mechanism and SCDown downsampling module were introduced to reduce the model complexity while ensuring detection accuracy. Experimental results indicate that the GNV2 network achieves excellent model compression effects while maintaining high performance: mAP@0.5 increased by 1.04%, reaching a total of 95.19%; model parameters were decreased by 41.95%, computational cost reduced by 36.71%, and the model size shrunk by 40.44%. Moreover, the GNV2-SLAM system incorporates point and line feature extraction techniques, effectively mitigate issues reduced feature point extraction caused by excessive dynamic targets or blurred images. Testing on the TUM dataset demonstrate that GNV2-SLAM significantly outperforms the traditional ORB-SLAM2 system in terms of positioning accuracy and robustness within dynamic environments. Specifically, there was a remarkable reduction of 96.13% in root mean square error (RMSE) for absolute trajectory error (ATE), alongside decreases of 88.36% and 86.19% for translation and rotation drift in relative pose error (RPE), respectively. In terms of tracking evaluation, GNV2-SLAM successfully completes the tracking processing of a single frame image within 30 ms, demonstrating expressive real-time performance and competitiveness. Following the deployment of this system on inspection robots and subsequent experimental trials conducted in the cowshed environment, the results indicate that when the robot operates at speeds of 0.4 m/s and 0.6 m/s, the pose trajectory output by GNV2-SLAM is more consistent with the robot's actual movement trajectory. This study systematically validated the system's significant advantages in target recognition and positioning accuracy through experimental verification, thereby providing a new technical solution for the comprehensive automation of cattle barn inspection tasks.