融合实例分割与视觉惯性里程计的丘陵山地果园语义SLAM 算法

Semantic SLAM Algorithm for Hilly and Mountainous Orchards Combining Instance Segmentation and Visual Inertial Odometry

摘要:

针对丘陵山地果园地势起伏导致的定位鲁棒性下降，以及传统SLAM因地形语义与高度信息缺失难以精确引导农机在丘陵山地果园下自主导航的问题，本文提出一种深度融合实例分割与视觉惯性里程计的语义SLAM算法。前端采用一种动态IMU预积分策略，在运动剧烈工况下引入龙格库塔（Runge-Kutta）法捕捉高频运动特征，增强位姿估计精度;同步利用集成CBAM与FPIoU-v2的改进YOLO v8-seg算法实现像素级语义提取，并通过语义投影最终构建具备高度、坡度、粗糙度及语义属性等多维地形信息的2.5D语义高程地图。试验结果表明，改进定位算法在不同丘陵果园实地测试环境下绝对位姿误差（APE）和均方根误差（RMSE）为0.38、0.24 m，优于VINS与ORB-SLAM3;改进YOLO v8-seg算法在自建数据集上mAP@0.5达到73.79%，推理速度为39.09 f/s，满足实时检测要求。本文所提算法可提升系统定位和感知能力，为果园履带车在非结构化环境下导航与作业提供理论与技术指导。

Abstract:

Aiming to address the issues of reduced positioning robustness caused by the steep topography of hilly and mountainous orchards, as well as the difficulty of accurately guiding agricultural machinery for autonomous navigation in such environments due to the lack of terrain semantics and elevation data in traditional SLAM, a semantic SLAM algorithm that deeply integrated instance segmentation with visual inertial odometry was proposed. In the frontend, a dynamic IMU pre-integration strategy was adopted, which introduced the Runge-Kutta method under intense motion conditions to capture high-frequency motion features, thereby enhancing the accuracy of pose estimation. Simultaneously, an improved YOLO v8-seg algorithm integrated with CBAM and FPIoU-v2 loss function was utilized for pixel-level semantic extraction. A 2.5D semantic elevation map containing multi-dimensional terrain information, including height, slope, roughness, and semantic attributes was finally constructed through semantic projection. Experimental results indicated that the root mean square error (RMSE) of the absolute pose error (APE) for the improved localization algorithm in different hilly orchard field tests was 0.38 m and 0.24 m, respectively, which outperformed VINS and ORB-SLAM3. The improved YOLO v8-seg algorithm achieved an mAP@0.5 of 73.79% on a self-built dataset with an inference speed of 39.09 f/s, satisfying real-time detection requirements. The proposed algorithm enhanced the localization and perception capabilities of the system, providing theoretical and technical guidance for the navigation and operation of tracked orchard vehicles in unstructured environments.