Learning-AI

RoR: Rules of the Road: Predicting Driving Behavior with a Convolutional Model of Semantic Interactions

April 2020

tl;dr: Multi-modal behavior prediction with perception output.

Overall impression

ChauffeurNet, IntentNet and Rules of the Road all uses semantic map that includes static and dynamic information.

For behavior prediction, we need past history of the agent, dynamics of other entities and semantic scene info. The paper does not specifically tackle the problem of motion planning. Rather it still focuses on the prediction of a target entity. This target entity could be ego car or other cars. does not explicitly perform prediction of other agents.

This is to be differed from the planning-centric ChauffeurNet. For motion planning, collision should be explicitly modeled.

The main novelty of this paper seems to be in the semantic map encoding into a 20-channel pseudo-map.

The rest of the paper is not well written. Lots of details are left out and the main body of the paper only have 8 pages for CVPR.

Key ideas

Technical details

Notes