An Integrating Comprehensive Trajectory Prediction with Risk Potential Field Method for Autonomous Driving

TL;DR

This work tackles safe and efficient autonomous driving under interaction uncertainty by integrating a region-based trajectory predictor (TRTP) with a risk potential field for planning. TRTP predicts trajectories to multiple target regions with associated probabilities, enabling comprehensive multi-modal forecasting; the risk potentials derived from these predictions are embedded into a Model Predictive Contouring Control (MPCC) planner to balance safety and progress. The approach achieves strong prediction coverage and risk-aware planning, validated in nuScenes-based training and CARLA simulations, with quantitative gains over baselines and qualitative demonstrations in left-turn and merging scenarios. Overall, the combination of region-based trajectory prediction and risk-aware optimization offers a practical path to non-conservative yet safe autonomous driving in interactive environments.

Abstract

Due to the uncertainty of traffic participants' intentions, generating safe but not overly cautious behavior in interactive driving scenarios remains a formidable challenge for autonomous driving. In this paper, we address this issue by combining a deep learning-based trajectory prediction model with risk potential field-based motion planning. In order to comprehensively predict the possible future trajectories of other vehicles, we propose a target-region based trajectory prediction model(TRTP) which considers every region a vehicle may arrive in the future. After that, we construct a risk potential field at each future time step based on the prediction results of TRTP, and integrate risk value to the objective function of Model Predictive Contouring Control(MPCC). This enables the uncertainty of other vehicles to be taken into account during the planning process. Balancing between risk and progress along the reference path can achieve both driving safety and efficiency at the same time. We also demonstrate the security and effectiveness performance of our method in the CARLA simulator.

Figures (7)

• Figure 1: Illustration of the risk potential fields and planning results at different time steps in the next 4 seconds. (a)At the current moment, the vehicle is preparing to merge into the traffic flow. (b) In the next 4 seconds, the risk potential field of each time step is established based on the results obtained by TRTP. The car in the risk potential field represents the pose of the ego vehicle at each time step in the future planned by the motion planning module mpcc.
• Figure 2: Illustration of TRTP. (a) shows that lanes are split into small pieces of equal length. Currently, the target vehicle is within the yellow lane piece, and after T seconds, it may be located in the red ones. Possible target paths are generated based on the topological connection relationships between lane pieces and are represented by blue arrows. (b) is the architecture of TRTP. TRTP consists of four modules. GRU encoder encodes the possible paths and historical trajectories of vehicles. Interactions encoder uses Scaled Dot-Product Attention to represent interactions between vehicles. The trajectory decoder and probability decoder decode a trajectory and corresponding probability for each path respectively.
• Figure 3: Kinematic bicycle model.
• Figure 4: $\theta _{r}$ is the projection of vehicle position onto the reference path. $\theta _{\mu}$ is an approximation of $\theta _{r}$ and the contouring error $e_{k}^{c}$ is approximated by $\hat{e}_{k}^{c}$, the lag error $e_{k}^{l}$ is approximated by $\hat{e}_{k}^{l}$.
• Figure 5: Qualitative results of our proposed prediction model. The pictures in the left column are HD maps of nuScenes in different scenes, The red box in a HD map represents the predicted target and the yellow boxes represent other vehicles. Different lanes are marked with lines of different colors, with white region indicating drivable areas and blue region indicating pedestrian zones. The pictures in the right column show the top 10 predicted trajectories with the highest probability and ground truth in different scenes.