A hybrid deep learning approach for driver anomalous lane changing identification

Area of Science:

• Road safety and intelligent transportation systems.
• Machine learning applications in behavioral analysis.
• Data science for driver behavior modeling.

Background:

• Understanding driving states is crucial for road safety.
• Anomaly detection in driving behavior identifies deviations caused by environmental or psychological factors.
• Personalized anomaly recognition is needed due to individual driving variations.

Purpose of the Study:

• To develop an efficient, unsupervised approach for personalized anomaly recognition of lane-changing events.
• To identify anomalous driving states in a manner tailored to individual drivers.
• To enhance road safety through better monitoring of driving behaviors.

Main Methods:

• Utilized a Recurrent-Convolutional Autoencoder for spatio-temporal feature extraction from high-dimensional naturalistic driving data.
• Employed Pauta criterion-based reconstruction loss analysis and one-class Support Vector Machine for personalized anomaly detection.
• Applied t-Distributed Stochastic Neighbor Embedding for latent space visualization and grey relational coefficient analysis for temporal anomaly identification.

Main Results:

• The proposed framework effectively captured personalized driving patterns and abnormal lane-changing events without prior labels.
• Results demonstrated heterogeneity in abnormal lane-changing behaviors across drivers and within individual temporal and spatial sequences.
• Validation on a dataset of 50 drivers and 59,372 lane change events confirmed the model's efficacy.

Conclusions:

• The unsupervised hybrid approach offers a novel method for personalized anomalous lane-changing behavior identification using naturalistic driving data.
• The approach enables the extraction of natural individual driving behavior patterns.
• Insights gained can improve personalized driving behavior monitoring systems for enhanced road safety.