Driver Fatigue Detection Based on BiLSTM-At and Multi-Source Information Fusion

In response to the frequent occurrence of traffic accidents caused by fatigued driving, this paper analyzes the problems existing in current fatigue detection methods, such as misjudgment, poor real-time performance, and the singularity of the types of datasets supporting related research. This paper proposes a multi-source information fusion driver fatigue detection method based BiLSTM-At. By synchronously collecting the driver's facial features and heart rate features, a multi-source information fusion driver fatigue dataset is constructed, and a feature-level fusion strategy is adopted to integrate the two types of heterogeneous data into a 9-dimensional feature vector. The proposed BiLSTM-At model captures the temporal dependency of the fatigue state through the bidirectional LSTM layer and combines the attention mechanism to dynamically allocate the weights of key features, enhancing the model's ability to focus on important information. The experiments are verified based on a self-built dataset. The results show that the accuracy of our method in detecting fatigue after fusing multi-source information reaches 97.2%, which is 2.9% higher than that of the single heart rate feature model, and 9.3% and 7.4% higher than those of the traditional LSTM and BiLSTM models, respectively. The ablation experiments further prove that the bidirectional temporal modeling and the attention mechanism contribute 7.4% and 6.2% to the accuracy improvement, respectively. The multi-source fusion strategy significantly improves the generalization ability and robustness of the model. The experimental results demonstrate that our model has a higher accuracy in the driver fatigue detection task and also verify the necessity of multi-source information fusion for fatigue detection.