TITAN: Future Forecast using Action Priors
Fujiry0 opened this issue · comments
Short summary: Propose TITAN(Trajectory Infer-ence using Targeted Action priors Network),
a new model that incorporates prior positions, actions, and context to forecast the future trajectory of agents and future ego-motion.
Details
- Method: Supervised deep neural networks
- Input:
- The bounding box of agent
- A sequence of image patches, obtained from the bounding box
- Ego-motion (acceleration and yaw rate of the ego-vehicle)
- Output:
- New interaction module
- Incorporates actions of individuals in addition to their locations
- Use multi-task loss
- Improve the performance of multi-label action recognition.
- Agent Importance Mechanism (AIM)
- Identify objects that are more relevant for ego-motion prediction
- Introduce TITAN dataset
- Consists of 700 labeled video-clips (with odometry) captured from a moving vehicle on highly interactive urban traffic scenes in Tokyo.
- 50 labels
- Vehicle states and actions, pedestrian age groups, and targeted pedestrian action attributes
- Ego-motion information from an IMU sensor.
Dataset: https://usa.honda-ri.com/titan