Early Lane Change Prediction for Automated Driving Systems Using Multi-Task Attention-Based Convolutional Neural Networks

Mozaffari, Sajjad; Arnold, Eduardo; Dianati, Mehrdad; Fallah, Saber · 2022 · Crossref

DOI: 10.1109/tiv.2022.3161785

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical safety challenge of predicting lane change (LC) maneuvers in automated driving systems. Existing methods often detect maneuvers only after they have begun or fail to estimate key timings, such as the time-to-lane-change (TTLC), which are essential for proactive decision-making. The authors propose a novel multi-task attention-based Convolutional Neural Network (CNN) that simultaneously predicts the likelihood of LC maneuvers and the TTLC. This approach aims to provide early warnings by analyzing the traffic context surrounding a target vehicle, rather than relying solely on the vehicle's recent motion. The method utilizes a Bird’s Eye View (BEV) representation of the driving environment, centered on the target vehicle, as input. This BEV includes vehicle bounding boxes, lane markings, and drivable areas, with a higher resolution in the lateral dimension to emphasize lateral motion. An attention-based CNN extracts spatio-temporal features from stacked BEV frames over a two-second observation window. A spatial attention mechanism divides the environment into four quadrants (front/back, left/right) and assigns weights to focus on the most relevant areas influencing the maneuver. These features are shared between two sub-networks: a classifier for maneuver type (left lane change, right lane change, or lane keeping) and a regressor for TTLC. The model is trained using two curriculum learning schemes that gradually increase the complexity of training samples by expanding the maximum included TTLC and adjusting the loss ratio between classification and regression tasks. The model was evaluated on a large-scale public trajectory dataset collected from German highways. The results demonstrate that the proposed method outperforms state-of-the-art LC prediction models, particularly in long-term prediction horizons. By predicting both the type and timing of maneuvers, the system provides more actionable information for automated vehicles to execute smooth and safe decelerations or other maneuvers. The spatial attention mechanism also enhances interpretability by highlighting which parts of the surrounding environment contribute most to the prediction. The significance of this work lies in its ability to improve the safety and comfort of automated driving systems by enabling earlier and more precise predictions of lane changes. The multi-task formulation and attention mechanism offer a robust solution for understanding complex traffic interactions, addressing limitations of previous models that struggled with long-horizon predictions and lacked interpretability. This approach supports more agile driving decisions in high-speed highway scenarios, potentially reducing accidents associated with unsafe lane changes.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• lane changing
• anticipation