Using neural networks to model restricted traffic zones for autonomous vehicle navigation

Abstract

The navigation of an AV in an environment may be planned based on a model of a restricted traffic zone in the environment. Information of the environment (e.g., a vector map, information of one or more objects, a temporal sequence of semantic grids, a query grid, etc.) may be input into a neural network. The neural network may include a CNN and a GNN. The map and tracks may be input into the GNN. The temporal sequence of semantic grids or query grid may be input into the CNN. The neural network may output edges of the restricted traffic zone. The neural network may output a grid of points representing locations in the environment and information indicating drivability of each respective point. The neural network may output one or more polylines dividing the environment into regions and information indicating whether the AV can drive to or in each respective region.