[Lane Change] How would surround vehicles move? A Unified Framework for Maneuver Classification and Motion Prediction, IEEE Trans. on. IV 2018

Summary.

1) Predict the best trajectories of the neighboring vehicles around the ego vehicles

2) Collect trajectories from 360 degrees surround video systems and Lidars and Radars

3) Hidden Markov Model (HMM) for maneuver classification, deterministic and probabilistic path prediction model, and vehicle interaction model are used for the final trajectories

 

1. Input

. past positions and velocityes of the neighboring vehicles and ego vehicle

 

2. Algorithm

. first calculate likelihood of each maneuver

. second generate trajectories corresponding to the maneuver

. third optimize based on collision assessment

 

 

 

 

2.1 Manuever classification

. define 10 maneuver classes

. a separate HMM is trained to classify one maneuver class

2.2 Path prediction

. deterministic : average of simple motion models (constant velocity, accel, accel and turn rate)

. probabilistic : variational Gaussian mixture model

. final prediction = 1/2 * (deterministic + probabilistic)

 

2.3 Vehicle interaction

. minimize the collision probabilities of the neighboring cars by the following optimization

 

3. Dataset

4. Results