Point cloud 3D object detection algorithm based on local information fusion
A three-dimensional object detection algorithm with a local information encoding module and a subsequent cross-fusion module was proposed aiming at the current lack of accurate spatial position information for three-dimensional object detection algorithms based on point clouds. Global features were efficiently encoded using 3D sparse convolution during the feature extraction phase. The local information encoding module leveraged the intrinsic information within the object's point cloud,constructing fine-grained semantic details. The information was reweighted to enhance the representation of local features through a self-attention mechanism. A cross-fusion module was introduced to facilitate interaction between local and global features,resulting in enhanced object detection features. The proposed method was validated using the KITTI and Waymo datasets. The average precision at IoU 0.7 for easy,moderate and hard tasks achieved 91.60％,82.53％,and 77.83％,respectively on the KITTI dataset. The average precision at IoU 0.7 reached 74.92％ on the Waymo dataset.
point cloudsparse convolutionlocal informationattention mechanismcross fusion
万方数据
