Detection of abnormal driving behaviour from autonomous vehicle’s view using Siamese and fully connected neural networks

In mixed traffic environments with autonomous vehicles (AVs) and human-driven vehicles (HVs), detecting abnormal driving behaviours of HVs is critical for safety and efficiency. This study proposes a Siamese fully connected neural network (Sia-FCNN) model to enhance AVs’ ability to identify HVs’ abnormal behaviours. First, model inputs, motion features of preceding vehicles, are extracted from AV onboard sensors, and a threshold for abnormal behaviour is derived from the historical kinematic data, with manual verification for calibration. Besides, the abnormality identification model incorporates a Siamese neural network framework with two FCNNs and a Lambda layer is established. The model chieves an average accuracy of 96.74% and an average F1-Score of 0.9420 on ApolloScape dataset (from China). Compared to the Long Short-Term Memory (LSTM) based model, the Sia-FCNN model exhibits superior convergence speed, fitting capability, and adaptability to small-sample learning. A cross-dataset validation using Boreas dataset (from Canada) confirms the robustness of proposed model.