GenAD: An innovative approach to autonomous driving perception

In the perception phase, the GenAD project uses a standard bird's-eye view processing method to extract agent-centric scene markers that incorporate map features. In order to model the high-level structured features of each agent, the project uses a variational autoencoder to map the ground truth trajectory into a Gaussian distributed latent space. This approach not only simplifies the difficulty of directly modeling multiple possible trajectories, but also can efficiently represent complex motion patterns in a lower-dimensional space.