Semantic segmentation-based point cloud intensity completion method and system

Abstract

The invention discloses a semantic segmentation-based point cloud intensity completion method and system. The method comprises the steps of synchronously acquiring an RGB image and point cloud data ofa road surface by a camera and a laser radar; performing spatial transformation on the point cloud data by utilizing the transformation matrix to generate a two-dimensional reflection intensity projection drawing and a two-dimensional depth projection drawing; performing reflection intensity completion on the RGB image and the two-dimensional reflection intensity projection drawing to obtain a single-channel reflection intensity projection drawing; performing depth completion on the RGB image and the two-dimensional depth projection drawing to obtain a single-channel depth projection drawing;performing coarse-grained completion processing on the RGB image, the single-channel reflection intensity projection drawing and the single-channel depth projection drawing to obtain a two-dimensional coarse-grained reflection intensity projection drawing; performing semantic segmentation processing on the RGB image and the two-dimensional depth projection drawing to obtain a plurality of regionsto be completed; and performing fine-grained reflection intensity completion on the two-dimensional coarse-grained reflection intensity projection drawing according to the to-be-completed regions toobtain a two-dimensional reflection intensity projection completion drawing.

Description

technical field [0001] The invention belongs to the field of unmanned driving, and in particular relates to a point cloud intensity complement method and system based on semantic segmentation. Background technique [0002] Lidar is the main tool for unmanned driving data collection. Lidar on the roof of unmanned vehicles generally has 16 / 32 / 64 / 128 line lidar. The cost of lidar increases with the number of lines, but less The lidar of the wire harness collects less point clouds, and the sparse point clouds are difficult to use for high-precision calculations. [0003] In addition, in actual acquisition, the energy of the lidar return beam is affected by distance and propagation medium. Long-distance objects cannot return the beam or the energy of the returned beam is lower than the detection threshold; in addition, due to the influence of microscopic particles in the atmosphere, meteorological conditions such as rain, snow and fog will cause the energy loss of the beam emitt...

AI Technical Summary

Problems solved by technology

Long-distance objects cannot return the beam or the energy of the returned beam is lower than the detection threshold; in addition, affected by the microscopic particles in the atmosphere, weather conditions such as rain, snow and fog will cause the energy loss of the beam emitted by the lidar, resulting in the object being unable to return the beam or return the energy of the beam below detection threshold

Both of these conditions will eventually result in the receiver not being able to receive the signal, the received signal being weakened or doped with noise

[0004] The current point cloud completion work mainly has the following limitations: (1) limited to the point cloud completion of local objects, especially the completion of indoor scenes or single objects; (2) limited to the completion of local dependencies, that is, based on phase (3) In the completion of large-scale overall scenes, the focus is on depth completion, while less research has been done on reflection intensity completion

Images