Open Access

ARTICLE

Lane-Changing Intention-Aware Vehicle Trajectory Prediction via Mutual Information-Guided Feature Selection and a Hybrid Convolutional Neural Network–Transformer Architecture

Huaran Zhou¹, Gaoteng Yuan², Ping Qiu^1,*

1 School of Internet of Things, Nanjing University of Posts and Telecommunications, Nanjing, China
2 Faculty of Computer and Software Engineering, Huai’an University, Huai’an, China

* Corresponding Author: Ping Qiu. Email:

Computer Modeling in Engineering & Sciences 2026, 148(3), 31 https://doi.org/10.32604/cmes.2026.086179

Received 25 May 2026; Accepted 26 August 2026; Issue published 28 September 2026

Abstract

Accurate vehicle trajectory prediction and lane-changing intention recognition are essential for autonomous driving and advanced driver-assistance systems, as they support motion planning, collision avoidance, and risk assessment. Existing deep learning methods often use high-dimensional trajectory variables without explicitly evaluating their relevance, which may introduce redundant information and reduce computational efficiency. In addition, local motion variations and long-range temporal dependencies are frequently modeled in isolation, although both are important for representing lane-changing behavior. To address these limitations, this paper proposes a joint intention-and-trajectory prediction framework that combines mutual-information-guided feature selection with a hybrid convolutional neural network (CNN)–Transformer architecture. Mutual information (MI) is used to rank target-vehicle, surrounding-vehicle, and road-related features according to their relevance to lane-changing intention, and a compact feature subset is retained. The CNN branch extracts short-term local motion patterns, whereas the Transformer branch captures long-range temporal dependencies. Their fused representation is optimized through an intention-classification head and a trajectory-regression head. Experiments on the Next Generation Simulation (NGSIM) and highD naturalistic highway trajectory datasets show that the proposed framework achieves better results than the in-house CNN, long short-term memory (LSTM), and Transformer baselines under the same evaluation protocol. It achieves intention-recognition accuracies of 95.81% and 92.16% on NGSIM and highD, respectively, while also reducing trajectory-prediction errors. The results indicate that feature relevance analysis and complementary local–global temporal modeling can improve the accuracy, efficiency, and interpretability of intention-aware vehicle trajectory prediction.

---

Keywords

---