After clicking “Watch Now” you will be prompted to login or join.
PredictionNet: Predicting the Future in Multi-Agent Environments for Autonomous Vehicle Applications
Alexey Kamenev, NVIDIA | Nikolai Smolyanskiy, NVIDIA
GTC 2020
Predicting the future trajectories of road agents is an import part of the planning&control stack in autonomous vehicles. Deep-learning approaches can be superior to classical methods in this domain, because neural networks can learn to use context and environment as a prior to improve prediction. We'll present PredictionNet — a deep neural network (DNN) that can be used for predicting future behavior/trajectories of road agents in autonomous-vehicle applications. Our DNN takes a rasterized top-down view of the world provided by the perception system and computes future predictions from past observations. We'll present its architecture, training-data collection process, and our training procedures. We'll also show video demos of live predictions on our self-driving car.