ISC-PRIME-V1

Introduction

We make accurate prediction of vehicle future motion by combining model- and learning-based methods, named PRISC-Net.

Framework of Method

The framework of PRISC-Net is as below:

The framework of Learning-based Path Target Predictor is as below:

The framework of Learning-based Trajectory Evaluator is as below:

Experiment

Evaluation Metric

• minimum Average displacement Error (minADE): the $l_2$ distance between the most possible trajectory among $k$ predicted trajectories and the ground-truth, averaged over all future time steps ($k=6$ in this paper).
• minimum Final Displacement Error (minFDE): the $l_2$ distance between the most possible trajectory among $k$ trajectories and the ground-truth at the final time step of prediction.
• Miss Rate (MR): the ratio of cases where the displacement between the predicted endpoint and the ground-truth endpoint exceeds the pre-defined threshold $\beta$ ($\beta=2.0 m$ in this paper).
• Traffic Rule Violation Rate (TRV): the ratio of scenarios where any predicted trajectory violates traffic rule or scene context constraints. Typical cases include entering non-reachable area, speeding and retrograding.
  • Entering non-reachable area is the case that any point of any predicted trajectory lies in the non-reachable area.
  • Speeding means that the speed of any point in any predicted trajectory is exceeds the speed limit.
  • Retrograding means the angle between the driving direction of any point of any predicted trajectory and the lane exceeds $\frac{\pi}{2}$.

Implementation Details

All models are trained on a NVIDIA TITAN V100 GPU with $12$ GB memory, and the implementation details for each stage are as follows:

Candidate Target Prediction

​ For candidate target sampling, two points are sampled every meter from lane centerlines. The number of hidden units is set to $64$ for all 3-layer MLPs. The overall target predictor is trained for $80$ epochs using Adam optimizer with the batch size and initial learning rate set to $128$ and $1\times 10^{-3}$, respectively.

Trajectory Generation

In our experiment, the coefficients in Eq. \ref{v_para}, \ref{a_para}, \ref{theta_para} are set as: $k_j=0.1$, $k_v=k_s=1$, $\alpha_1=5$, $\alpha_2=2$ and $\alpha_3=\frac{\pi}{6}$.

Trajectory Evaluation

The INTERACTION dataset provides observed state sequence with a time interval of $\triangle T = 0.1s$, and the continuous trajectories are discretized with the same time interval. All reachable path inputs are discretized with a distance interval of $\triangle D = 2m$. We train the evaluator for $80$ epochs with the batch size and initial learning rate set to $128$ and $1\times 10^{-3}$, respectively. The evaluator is optimized with Adam with a decay of $10$ every $10$ epoch.

Expriment Result

Comparison Video