Method for generating a representation of the surroundings

Abstract

The present method generates a three-dimensional representation of surroundings using an end-to-end convolutional neural network. A captured image is subdivided into multiple vertical columns, which the model processes to parameterize multiple three-dimensional stixels for each column. By providing at least two stixels per column, the system captures a comprehensive environment, including elevated objects like barriers or tunnels. The stixels may be designed as slanted stixels, defined by bottom and top points with associated depth information and semantic classes. This efficient architecture enables high-resolution environmental mapping suitable for autonomous vehicle navigation on embedded hardware, effectively reducing the reliance on separate, computationally intensive depth estimation and semantic segmentation processes.