Point Cloud Registration Network Based on Daul-Attention Mechanism For Truck Pose Estimation

Truck pose estimation is a pivotal technology for automating excavation and loading processes in intelligent mining excavators. The existing methods for pose estimation are mainly based on sensor measurements and traditional registration techniques. While these sensor-based methods can identify truck poses, they are prone to significant cumulative errors due to factors like vibrations and also suffer from short sensor lifespans. Traditional registration methods use iterative closest point algorithms to estimate the bucket’s pose by aligning its point clouds, but these methods are prone to getting stuck in local optima. Recently, deep learning-based networks for object pose estimation have been introduced. However, in the mining operation environment, the truck’s position changes significantly during the loading process, leading to distinct truck features at different poses. Traditional point cloud registration networks often just concatenate the features of template and target point clouds, which inadequately exploits both global and local features. This limitation decreases the accuracy of truck pose estimation. To tackle the feature fusion challenge, we introduce a novel feature fusion registration network named CASNet. CASNet consists of a feature extraction module, a feature fusion module, and a pose computation module. The feature fusion module employs spatial and channel attention mechanisms to effectively learn different hierarchical features during feature fusion. By integrating both local and global features, CASNet addresses feature fusion issues and enhances generalization capabilities. Experimental results show that our method significantly reduces both rotational and translational errors in truck pose estimation by 1.26° and 2.74 cm, respectively.