基于YOLO模型的目标位姿估计

摘要:

本文针对共轴双旋翼无人机抓取应用中的目标位姿估计方法展开研究，提出基于CDPN的无人机目标位姿估计模型EPRO-CDPN，其前置目标检测器采用改进的YOLO网络算法，提高其目标检测能力；引入注意力机制，使模型关注关键特征信息，加强网络训练过程中通道间的特征融合；引入EPRO-PnP替换原来的传统PnP方法，将传统求解转化为对位姿概率分布的预测。整个位姿估计网络实现为一个端到端的网络。位姿估计网络模型在在公开数据集LineMod、自制数据集上进行了算法性能测试，并以共轴双旋翼无人机为物体目标进行抓取实验，验证姿态估计算法的可行性和有效性。检测精度在95%以上，检测速度快，FPS达到35.2帧/秒，实现了实时目标姿态估计，为视觉引导的机械臂自动抓取回收共轴无人机奠定了研究和实验方案。

Abstract:

This article investigates target pose estimation methods for co-axial dual-rotor drone grasping applications, proposing the EPRO-CDPN model for drone target pose estimation based on CDPN. The front-end target detector employs an improved YOLO network algorithm to enhance its target detection capability. An attention mechanism is introduced to enable the model to focus on critical feature information, strengthening the feature fusion between channels during the network training process. The traditional PnP method is replaced with EPRO-PnP, transforming the conventional solving process into a prediction of pose probability distribution. The entire pose estimation network is implemented as an end-to-end network. The pose estimation network model has undergone performance testing on the public dataset LineMod and a self-created dataset, conducting grasping experiments with the co-axial dual-rotor drone as the target object to verify the feasibility and effectiveness of the pose estimation algorithm. The detection accuracy exceeds 95%, with a fast detection speed achieving 35.2 frames per second, enabling real-time target pose estimation and laying the groundwork for research and experimental schemes for visually guided robotic arm automatic grasping and recovery of the co-axial drone.