Research map / Label-efficient and robust learning / Recognition and long-tail

Dual-View Interaction-Aware Lane Change Prediction for Autonomous Driving

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a dual-view interaction-aware model that fuses perceived-safety future interaction features (TTC/TET/TIT/SRP risk metrics with a Graph Transformer) and historical interaction features for lane change intention prediction on NGSIM and HighD.

Problem

Lane change intention prediction for surrounding human-driven vehicles in mixed-autonomy traffic; existing interaction-aware methods model only historical interactions and ignore future interactions, leading to unsafe predictions.

Approach

Adds a Future Interaction Module that models perceived safety via four spatio-temporal risk metrics (ameliorated TTC, TET, TIT and a spatial Moran's-I-style index SRP), combined with a Historical Interaction Module (Graph Transformer over vehicle states) and an Interaction Fusion Module (contrastive regularization and edge-biased self-attention).

Key results

• Achieves average 11.7-12.4% classification improvement over best baselines on NGSIM and HighD (e.g., 97.62%/98.33% left-lane-change precision/recall on NGSIM vs. 88.46%/86.38% for RFormer).
• Forecast ability (RHCF) improves on average 75.6-95.7% over best baselines across the two datasets.
• Removing the Future Interaction Module drops classification performance by about 10.5% on both datasets.

Measured effects reported in the paper

Surprising findings

• On 100 manually labeled safety-critical samples per dataset, the proposed model reaches 78% (NGSIM) and 74% (HighD) accuracy versus 16-17% and 10-11% for the best baselines.

Limitations

• Future interaction modeling relies on hand-designed risk metrics with fixed thresholds (e.g., TTC threshold 2.5s, tau 0.1s).
• Evaluation limited to two highway datasets with manually labeled safety-critical subsets of 100 samples each.

Experimental setup

Datasets: NGSIM HighD

Benchmarks: NGSIM HighD

Metrics: Precision Recall RHCF

Baselines: RF LSTM Bi-LSTM TMMOE MMAE VWC H-LSTM P-GHMM RFormer

Builds on: Graph Attention Networks Transformer Moran's I

Closest papers by meaning

• BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving (AAAI 2024 · similarity 0.86)
• LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.86)
• Hybrid-Driving: An Autonomous Driving Decision Framework Integrating Large Language Models, Knowledge Graphs and Driving Rules (AAAI 2025 · similarity 0.85)
• PSI: A Benchmark for Human Interpretation and Response in Traffic Interactions (NeurIPS 2025 · similarity 0.85)
• Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.84)
• Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 2025 · similarity 0.84)