Sparse Bird’s-Eye-View Backbone for Autonomous Vehicle Perception

Overview

Perception in Autonomous Vehicles Using a Sparse Bird's-Eye-View Backbone Neural Network is a novel sparse architecture I developed in collaboration with Waymo for efficient perception in self-driving vehicles. The work is patent pending.

The backbone leverages the inherent sparsity of bird's-eye-view (BEV) representations to dramatically reduce computation while maintaining perception quality. Building on this architecture, I trained a sparse multi-sensor fusion model for lane line, obstacle, and vehicle prediction that is now in production at Waymo.