Point cloud feature extraction method, device and equipment and storage medium

Abstract

The invention discloses a point cloud feature extraction method, device and equipment and a storage medium. The method comprises: dividing the space of collected point cloud data into a plurality of voxel grids, and determining the height of the voxel grids according to the space height of the point cloud data; performing convolution and pooling operation on the point cloud data in each voxel gridto obtain a global feature vector layer of the point cloud data; calculating a height feature vector layer and a density feature vector layer of the point cloud data; and combining the global featurevector layer, the height feature vector layer and the density feature vector layer to obtain a combined feature map of the point cloud data. According to the technical scheme, the global feature vector, the density feature vector and the height feature vector are combined, rich point cloud features are extracted, the voxel grids are divided according to the spatial height of the point cloud data,the number of the voxel grids is reduced, the calculation amount is effectively controlled, and the feature extraction efficiency is considered while the feature expression richness of the feature map is improved.

Description

technical field [0001] Embodiments of the present invention relate to the technical field of computer vision, and in particular to a point cloud feature extraction method, device, equipment, and storage medium. Background technique [0002] With the maturity of deep learning technology, the related research of intelligent vehicles, such as the development of peripheral sensing components, positioning and mapping, planning and decision-making, and environmental perception, have made certain breakthroughs. As a core technology, environmental perception usually involves multi-sensor fusion. Among them, lidar plays an important role in environmental perception technology because of its accuracy and reliability in positioning and ranging. Lidar can truly reflect the position and shape of the target by acquiring sparse point cloud data, but lacks rich texture information in the image, and has high requirements for feature extraction of point cloud data. [0003] The current featu...

AI Technical Summary

Problems solved by technology

[0003] The current feature extraction method for point cloud data is usually to discretize the point cloud into a multi-channel BEV (bird's eye view), and store features such as point cloud height or density values ​​in each layer of BEV, which is limited by the number of layers and resolution , this method is not rich enough for feature extraction, and will cause a certain degree of feature loss; there are also some methods that divide the point cloud into voxel grids in order to improve the object detection effect, and use convolutional networks to The feature extraction of quantitative point clouds can improve the richness of feature extraction to a certain extent, but it will lead to increased calculation and low feature extraction efficiency.

Images