基于多传感融合的农业动态复杂场景SLAM算法

Agricultural Dynamic Complex Scene SLAM Algorithm Based on Multi-sensor Fusion

摘要:

针对农业复杂动态环境中存在的强光照干扰、重复纹理结构及移动物体干扰等问题导致传统单传感器建图精度不足，提出一种设施大棚内多传感器融合SLAM算法。首先构建基于YOLO动态目标检测的高鲁棒性视觉SLAM算法（Visual SLAM algorithm based on YOLO dynamic object detection，YDOD-VSLAM），在ORB-SLAM3框架中引入改进YOLO v8n模型，使用MobileViT替换原主干网络以增强泛化能力，并在检测头前加入SimAM注意力机制提升遮挡物体识别性能。由于ORB-SLAM3的特征提取能力有限且仅能构建稀疏地图，引入目标检测线程识别动态特征点并结合前端过滤机制剔除动态干扰，增加稠密地图和八叉树地图模块实现对环境三维几何结构与语义信息的融合表达。基于A-LOAM算法构建激光局部栅格地图，通过贝叶斯估计将其与YDOD-VSLAM生成的视觉栅格地图融合形成全局栅格地图。在TUM数据集试验结果表明，YDOD-VSLAM在高动态数据集的绝对轨迹误差较ORB-SLAM3降低89.9%，较DynaSLAM降低36.5%。大棚构建地图试验结果表明，多传感器融合地图平均绝对误差为0.11 m，维持在较低水平范围，相较于YDOD-VSLAM平均误差降低52%，完整性与地图定位精度都优于单传感器建图，能为农业机械定位与导航提供更可靠的环境信息。

Abstract:

In view of the problems such as strong light interference, repetitive texture structures and moving object interference existing in the complex and dynamic environment of agriculture, which lead to the insufficient accuracy of traditional single-sensor mapping, a multi-sensor fusion SLAM algorithm for facilities greenhouses was proposed. Firstly, a highly robust visual SLAM algorithm based on YOLO dynamic object detection (Visual SLAM algorithm based on YOLO dynamic object detection, YDOD-VSLAM) was constructed. The improved YOLO v8n model was introduced into the ORB-SLAM3 framework, and the original backbone network was replaced with MobileViT to enhance the generalization ability. Additionally, a SimAM attention mechanism was added before the detection head to improve the recognition performance of occluded objects. Due to the limited feature extraction capability of ORB-SLAM3 and its inability to construct a dense map, a target detection thread was introduced to identify dynamic feature points and combine the front-end filtering mechanism to eliminate dynamic interference. The dense map and Octree map modules were added to achieve the fusion expression of the three-dimensional geometric structure and semantic information of the environment. Based on the A-LOAM algorithm, a laser local grid map was constructed, and it was fused with the visual grid map generated by YDOD-VSLAM through Bayesian estimation to form a global grid map. The experiments on the TUM dataset showed that the absolute trajectory error of YDOD-VSLAM in high dynamic datasets was 89.9% lower than that of ORB-SLAM3 and 36.5% lower than that of DynaSLAM. The mapping experiment conducted in the greenhouse showed that the average absolute error of the multi-sensor fusion map was 0.11 m, remaining within a relatively low range. Compared with the average error of YDOD-VSLAM, it was reduced by 52%. Both the completeness and the map positioning accuracy were superior to those of single-sensor mapping, providing more reliable environmental information for the positioning and navigation of agricultural machinery.