Learning-AI

TrafficPredict: Trajectory Prediction for Heterogeneous Traffic-Agents

June 2019

tl;dr: Propose a 4D graph representation of trajectory prediction problem. The paper also introduced a new dataset in ApolloScape.

Overall impression

The prediction seems not that good? Admittedly the dataset is focused on urban driving scenario and the problem is much harder than highway driving scenarios. The predicted trajectories projected on the single camera frame do not quite make sense either (some agents goes above the horizon and goes to the sky?)..

Key ideas

Technical details

Notes