🤖 AI

A Dynamic Scene Interaction Reasoning Framework for Scene-level Lane-Change Intention and Trajectory Prediction of Multiple Interacting Vehicles

This paper proposes DSiGAT, a dynamic scene graph attention framework that models traffic as a time-varying interaction graph to simultaneously predict lane-change intentions and future trajectories for multiple vehicles, achieving superior accuracy and scene-level consistency compared to existing baselines on standard datasets.

Original authors: Joshua Kofi Asamoah, Blessing Agyei Kyem, Eugene Denteh, Armstrong Aboah

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Joshua Kofi Asamoah, Blessing Agyei Kyem, Eugene Denteh, Armstrong Aboah

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a busy highway from a drone. Most traffic prediction systems today are like a single-minded detective who only watches one specific car, trying to guess where that car is going, while treating everyone else as just blurry background noise. They might say, "That red car is changing lanes," but they don't really know what the blue car behind it will do, or how the green car in front will react.

This paper introduces a new system called DSiGAT (Dynamic Scene Interaction Graph Attention Network) that acts more like a conductor of a massive orchestra. Instead of focusing on just one instrument, it listens to the entire band at once to understand how they are all playing together.

The Big Idea: The Traffic Dance Floor
The authors argue that predicting traffic isn't about guessing one car's next move in isolation. It's about understanding a "dynamic scene." They built a system that treats every car in a local group as a dancer on a floor.

The Graph: Imagine the cars are nodes (dots) and the invisible lines of influence between them are edges. If a car is close to another, a line connects them.
The Attention: The system uses "graph attention" to figure out which dancers matter most at any given second. It's like the conductor noticing that the violinist is about to speed up, so the drummer needs to adjust immediately.
The Goal: The system doesn't just guess where a car will be in 4 seconds; it guesses what the car intends to do (stay in lane, move left, or move right) and then uses that intention to draw the future path.

What They Rejected (The "No-Go" Zone)
The paper explicitly argues against the old-school method of "target-centered" prediction. They say that picking one "target" car and ignoring the future behaviors of everyone else is a dead end. If you only predict one car, you miss the ripple effect. For example, if Car A moves left, Car B might have to slow down. A system that only predicts Car A won't know Car B is about to brake, leading to a confused and potentially unsafe prediction. The authors show that looking at the whole scene together is necessary to get the physics right.

The Results: How Well Did It Work?
The team tested their "orchestra conductor" on real highway data from the NGSIM I-80, NGSIM US-101, and highD datasets. Here is what they found:

Guessing Intentions: The system was incredibly good at guessing what drivers wanted to do. On the I-80 highway, it got the right answer 90.12% of the time. On US-101, it hit 90.97%. This is a big jump compared to previous methods.
Predicting the Path: When it came to drawing the future path of the cars, the system was much more accurate than the best previous models. It reduced the error (measured as Root Mean Squared Error, or RMSE) by up to 52.94% on the highD dataset.
Avoiding Crashes: Because the system predicts everyone at once, it ensures the predictions don't crash into each other. It lowered the "Inter-Agent Collision Rate" (how often the predicted paths would hit) significantly compared to other models.
Early Warning: The system is good at spotting a lane change early. Even 3 seconds before a driver actually starts moving, the system can often tell what they are planning to do.

The "Secret Sauce" (How They Did It)
The authors didn't just throw a lot of data at a computer; they built specific tools to make it work:

Dynamic Graphs: They didn't use a fixed list of neighbors. The connections between cars change as they speed up, slow down, or switch lanes.
Intention-Guided Decoding: The system first guesses the intention (e.g., "Left Lane Change"), and then uses that guess to guide the drawing of the path. It's like deciding "I'm going to the kitchen" before you start walking, rather than just wandering and hoping you end up there.
Scene Consistency: They added a rule during training that says, "Hey, if you predict Car A moves left, make sure Car B doesn't predict it will move into the exact same spot." This forces the whole scene to make sense together.

How Sure Are They?
The authors are very confident in these numbers because they are based on measured results from real-world driving data, not just simulations. They tested their model against many other famous models (like Social LSTM, Graph Neural Networks, and Transformers) and consistently beat them.

They also checked how robust the system is. When they added "noise" (simulating bad sensors or messy data) to the input, the system didn't crash; it just got slightly less accurate, which is exactly what you want in a real-world safety system.

The Bottom Line
This paper suggests that to predict traffic safely, we need to stop looking at cars as isolated islands. By treating a group of cars as a single, interacting system where everyone's move affects everyone else, we can predict intentions and paths with much higher accuracy. The system, DSiGAT, proved it can do this better than current methods, reducing errors by up to half and keeping the predicted future of the road collision-free. It's a step toward self-driving cars that don't just see the car in front of them, but understand the whole dance of the highway.

Technical Summary: A Dynamic Scene Interaction Reasoning Framework for Scene-level Lane-Change Intention and Trajectory Prediction

Problem Statement
Safe motion planning for Advanced Driver-Assistance Systems (ADAS) and Autonomous Vehicles (AVs) requires an accurate understanding of how the surrounding traffic scene evolves. Existing lane-change prediction methods often suffer from two primary limitations:

Target-Centered Formulation: Many approaches focus on a single designated target vehicle, treating surrounding vehicles merely as passive context. This fails to capture how multiple vehicles evolve together and influence one another within a local scene.
Implicit Maneuver Representation: While multi-agent trajectory forecasting has advanced, it frequently describes scene evolution solely through future positions. These methods often lack explicit characterization of lane-change behaviors (e.g., left vs. right change vs. lane keeping), making it difficult to reason about the semantic meaning of the forecasted motion.

The core challenge identified is to represent the future traffic scene in a form that explicitly captures both the lane-change intention and the corresponding motion of every relevant vehicle while ensuring the predictions remain mutually compatible.

Methodology: DSiGAT Framework
The paper proposes the Dynamic Scene Interaction Graph Attention Network (DSiGAT), a framework designed for scene-level joint lane-change intention and trajectory prediction. The methodology proceeds through the following stages:

Dynamic Scene Graph Representation:
- The local traffic scene is modeled as a time-varying interaction graph where vehicles are nodes and their relationships are directed edges.
- Nodes: Represent vehicle states including longitudinal/lateral speed, acceleration, displacement, and one-hot lane encoding.
- Edges: Directed edges connect vehicles in the same or adjacent lanes within a 100m longitudinal gap. Edge features encode pairwise relationships, including relative gaps, lateral offsets, relative speeds, and a Time-to-Collision (TTC) metric.
- Temporal Sequence: The observation window (3 seconds, 30 frames) is represented as a sequence of graphs, capturing the evolution of interactions.
Temporal Graph Attention Message Passing:
- The framework employs a Graph Attention Network (GAT) to propagate interaction information. It computes attention scores based on node features, edge features, and the target node's state.
- Temporal Aggregation: A temporal attention mechanism weights the graph representations across the observation window, allowing the model to focus on the most informative frames and interaction cues preceding a maneuver.
Intention-Guided Trajectory Decoding:
- Intention Head: A feedforward network predicts a probability distribution over three maneuver classes: Lane Keeping (LK), Left Lane Change (LLC), and Right Lane Change (RLC).
- Decoupled Decoders: Instead of a single decoder, the framework utilizes three maneuver-specific trajectory decoders (one for each class).
- Weighted Mixture: The final predicted trajectory is a weighted combination of the three decoder outputs, where the weights are the predicted intention probabilities. This ensures semantic alignment between the predicted maneuver and the future path.
Training Objectives and Constraints:
- The model is trained end-to-end using a composite loss function:
  - Intention Loss ( $L_{int}$ ): Cross-entropy for maneuver classification.
  - Trajectory Loss ( $L_{traj}$ ): Mean Squared Error (MSE) for trajectory regression.
  - Intention-Guided Consistency ( $L_{gate}$ ): Penalizes mismatches between the predicted maneuver class and the lateral direction of the predicted trajectory.
  - Scene-Level Consistency ( $L_{scene}$ ): A physical constraint that penalizes predicted futures where vehicles violate a minimum separation distance, discouraging mutually incompatible trajectories.

Key Contributions
The paper outlines four main contributions:

Scene-Level Multi-Agent Formulation: Lane-change prediction is reformulated as a unified problem where the intention and trajectory of all valid vehicles in a local scene are predicted simultaneously, rather than focusing on a single target.
Dynamic Scene Interaction Graph: A novel graph construction that encodes time-varying spatial and motion relationships via directed edges and graph attention, moving beyond fixed neighbor slots.
Intention-Guided Decoding: A mechanism that explicitly couples maneuver prediction with motion forecasting, using the intention distribution to guide the trajectory decoder.
Scene-Level Consistency Constraint: A training objective that enforces physical plausibility across the entire scene, reducing inter-agent collision rates and joint displacement errors.

Experimental Results
The framework was evaluated on the NGSIM (I-80 and US-101) and highD datasets.

Intention Prediction: DSiGAT achieved state-of-the-art accuracy, reaching 90.12% on NGSIM I-80 and 90.97% on US-101. It outperformed strong baselines (including Dual Transformer and MCLG) across all maneuver classes (LK, LLC, RLC) and demonstrated robust early anticipation performance up to 3 seconds before maneuver onset.
Trajectory Prediction: The model significantly reduced trajectory Root Mean Squared Error (RMSE). On NGSIM I-80, it reduced RMSE by up to 52.94% relative to the strongest baseline (GRIP++). Similar improvements were observed on US-101 and highD.
Scene-Level Consistency: DSiGAT achieved the lowest Inter-Agent Collision Rate (IACR) and Joint Displacement Error (JDE) among all compared methods, indicating that its predictions are not only accurate individually but also coherent as a group.
Efficiency: Despite the complex scene-level reasoning, the model remains computationally efficient with 1.02 million parameters and an inference time of 167.68 ms per scene on an NVIDIA A100 GPU, which is faster than several Transformer-based baselines.

Significance and Claims
The authors claim that DSiGAT addresses the gap between intention prediction and multi-agent trajectory forecasting by providing a unified, behavior-aware scene representation. The significance of the work lies in its ability to:

Move beyond target-centered prediction to model the coupled futures of interacting vehicles.
Explicitly link maneuver semantics (intention) with motion forecasting, improving the interpretability and reliability of predictions.
Generate physically coherent scene-level forecasts that reduce the risk of conflicting predictions (e.g., collisions) in mixed traffic.

The paper concludes that the proposed dynamic graph approach, combined with intention-guided decoding and scene-level consistency constraints, offers a more robust and reliable framework for anticipating traffic evolution in mixed-autonomy environments.

More like this