Vehicle lane-change intention recognition based on BiLSTM Attention model for the Internet of vehicles

In terms of lane-changing and other driver actions, precise identification of the intentions of nearby vehicles is crucial to autonomous vehicle performance safety. At present, research in this domain primarily focuses on ideal environments without considering data packet loss. Therefore, this paper considered the impact of packet loss in the Internet of Vehicles on the performance of the lane change intent recognition model. To achieve this, an enhanced BiLSTM Attention model, which combines the bidirectional long short-term memory network structure and attention mechanism, is proposed based on LSTM. The NGSIM (Next Generation Simulation) dataset was utilized to extract vehicle lane-change behaviors for model training and testing. A long short-term memory (LSTM) model was employed to conduct comparative experiments using various input frequencies and packet loss rates. The performance of the proposed BiLSTM Attention model was evaluated through ablation experiments. A comparison was made between the model's performance in the absence of packet loss and its performance under a scenario with 30% packet loss. Additionally, the impact of continuous packet loss on the recognition of the lane-change intent model was analyzed. Experiments show that it outperforms basic LSTM and BiLSTM models, including the LSTM Attention method, with impressive improvements of 7.84%, 2.22%, and 4.89% (F1macro) and 2.83%, 1.03%, and 2.18% for the area under the receiver operating characteristic curve (AUC), respectively. Even under extreme (30%) packet loss conditions, the proposed model outperforms the same models by 8.23%, 2.68%, and 5.38% (F1macro) and 2.94%, 1.03%, and 2.29% (AUC), respectively. For 30% packet loss, the proposed model's performance decreased by 0.108% (F1macro) and 0.102% (AUC); however, the LSTM, BiLSTM, and LSTM Attention model performances decreased by 0.468% and 0.209%, 0.554% and 0.103%, and 0.569% and 0.208% for F1macro and AUC, respectively. Thus, the proposed model is the least affected by packet loss.