An unmanned flow vehicle drivable area perception method and system based on a lightweight semantic occupancy grid

Abstract

The application relates to the technical field of automatic driving perception, in particular to a method and system for perceiving drivable areas of an unmanned logistics vehicle based on a light-weight semantic occupancy grid, which fuses a surround view image and a laser radar point cloud, extracts BEV grid features, and predicts an occupancy probability; a semantic label is identified for a high-occupancy grid to generate a preliminary feasible area; a time sequence is smoothed in combination with a previous frame result to output a final feasible area of a current frame, thereby realizing the perception of drivable areas of the unmanned logistics vehicle. The application can realize occupancy grid perception only by using a low-line laser radar and a camera, greatly reduces hardware cost, and meets the commercialization requirements of the unmanned logistics vehicle. A sparse convolution and a dynamic activation area strategy are adopted, and the calculation amount is only 20%-30% of that of a dense method, so that the application is suitable for edge deployment. The application does not depend on preset categories, can detect various unknown obstacles, and improves driving safety. Semantic information is added on the basis of occupancy, richer environment understanding is provided for a planning module, and the decision-making quality is optimized.

Images