BiLSTM and Transformer Neural Network for Abnormal Driving Behavior Recognition

Along with the escalating demand for traffic supervision and the swift development of assisted driving, there are heightened requirements for the accuracy and diversity of identifying abnormal driving behaviors. Existing research is predominantly contingent upon trajectory data, resulting in limited types of abnormalities recognized in activities such as lane changes and turns. In response to this limitation, this paper proposes a method for classifying driving behaviors by integrating environmental factors and employing an identification model based on Bidirectional Long Short-Term Memory (BiLSTM) networks and Transformer. By incorporating road environment data to precisely define driving behaviors and expanding the range of abnormalities, the model utilizes BiLSTM to extract bidirectional features of driving behaviors in time series. Furthermore, in tandem with the Transformer, it can capture global information for feature fusion, culminating in the identification of abnormal driving behaviors. Experimental results indicate that the recognition accuracy can reach 98.85%, demonstrating an improvement of 0.4% when juxtaposed with models like ResNet.