A temporal information-based network model for vehicle lane-change behavior recognition

作者：Yafei Liu, Mujiao Ouyang, Xiaoguo Zhang · 年份：2025 · DOI：10.1117/12.3070249 · 研究领域：Traffic Prediction and Management Techniques、Neural Networks and Applications、Anomaly Detection Techniques and Applications

Accurate recognition of vehicle lane-change behavior is crucial for autonomous driving. Traditional approaches relying on IMU data or visual information are often constrained by inherent sensor errors and environmental obstructions, leading to limited performance. Moreover, the scarcity of publicly available datasets has impeded the development of robust lane-change recognition systems. To address these challenges, this study proposes a low-cost, efficient, and accurate method for vehicle lane-change behavior recognition, applicable to lane-level vehicle localization. A dedicated dataset was meticulously constructed, integrating multiple inputs—acceleration, angular velocity, and lateral distance— to mitigate uncertainty errors from single-sensor data, capturing critical parameters of lane-change behavior. Leveraging the temporal nature of lane-change actions, a hybrid CNN+TCN+BiLSTM network model was developed. The Temporal Convolutional Network (TCN) residual block enhances short- and long-term temporal dependency modeling, while the BiLSTM captures bidirectional sequential patterns, and a 3D tensor-based attention mechanism extracts key time-step information from the data. The proposed model was rigorously trained and tested on the constructed dataset, achieving a recognition accuracy of 99.5%. Real-world road tests confirmed its robustness in complex urban environments, maintaining reliable performance even under suboptimal visual conditions. This method effectively ident...