KI-PMF: Knowledge Integrated Plausible Motion Forecasting

TL;DR

KI-PMF addresses safe, multimodal motion forecasting for autonomous driving by marrying explicit knowledge priors with learning. A deterministic non-parametric refinement layer prunes infeasible trajectories using HD-map constraints and kinematics, followed by a learnable encoder that fuses refined trajectories with goal lanes through cross-attention to produce a probability distribution over feasible futures. Key contributions include the two-stage design, explicit reachability guarantees, lane–trajectory interaction modeling, and real-time capable pruning, all validated on Argoverse with high drivable-area compliance. The approach enhances safety and reliability in real-world autonomous driving by reducing off-road and physically implausible predictions while maintaining competitive predictive accuracy.

Abstract

Accurately forecasting the motion of traffic actors is crucial for the deployment of autonomous vehicles at a large scale. Current trajectory forecasting approaches primarily concentrate on optimizing a loss function with a specific metric, which can result in predictions that do not adhere to physical laws or violate external constraints. Our objective is to incorporate explicit knowledge priors that allow a network to forecast future trajectories in compliance with both the kinematic constraints of a vehicle and the geometry of the driving environment. To achieve this, we introduce a non-parametric pruning layer and attention layers to integrate the defined knowledge priors. Our proposed method is designed to ensure reachability guarantees for traffic actors in both complex and dynamic situations. By conditioning the network to follow physical laws, we can obtain accurate and safe predictions, essential for maintaining autonomous vehicles' safety and efficiency in real-world settings.In summary, this paper presents concepts that prevent off-road predictions for safe and reliable motion forecasting by incorporating knowledge priors into the training process.

Figures (4)

• Figure 1: The left column represents the two scenes with unrefined trajectories injected into the scene. The right column represents the impact of our refinement layer, where we refine the original trajectory set with priors, allowing us to avoid predicting in the off-road areas.
• Figure 2: The green area delineates the local buffer polygon, instrumental in diminishing the spatial extent of the principal polygonal; non-drivable regions. The blue linear entities correspond to the polygonal boundaries. For illustrative purposes, a circular buffer, depicted by a red dotted circumference, is employed. Subsequently, the peripheries of the green polygons undergo collision detection analysis to expedite the trajectory computation within the polygonal confines The x and y-axis are represented in the city coordinates system.
• Figure 3: High-level overview of the architecture. The Refinement layer takes two inputs. 1. Lane centerline points in the map coordinate system for a given query window (in meters) and 2. Trajectory sets computed from the dataset with a $\epsilon = 2$ coverage. The refinement layer produces feasible/pruned trajectories by constructing a lane boundary given the lane-centerline points. It also produces a list of possible goal positions or reachable lanes where the target actor could reach within a given prediction horizon. LaneGCN is used as a backbone architecture and outputs an embedding representation for $a_{tar}$ which is then concatenated with the lower parts of the network.
• Figure 4: Performance of various methods to employ Point-in-Polygon operation. We use $Ours\_RayTracing\_Parallel$ algorithm in our refinement layer for pruning non-compliant trajectories.