Deep learning approach for unified recognition of driver speed and lateral intentions using naturalistic driving data

Driver intention recognition is a critical component of advanced driver assistance systems, with significant implications for improving vehicle safety, intelligence, and fuel economy. However, previous research on driver intention recognition has not fully considered the influence of the driving environment on speed intentions and has not exploited the temporal dependency inherent in the lateral intentions to prevent erroneous changes in recognition. Furthermore, the coupling of speed and lateral intentions was overlooked; they were generally considered separately. To address these limitations, a unified recognition approach for speed and lateral intentions based on deep learning is presented in this study. First, extensive naturalistic driving data are collected, and information related to road slope and driving trajectories is extracted. A comprehensive classification of driver intentions is then performed. Toeplitz inverse covariance-based clustering and trajectory clustering methods are applied separately to label speed and lateral intentions, so that the influence of driving environments and the coupling of speed and lateral intentions are integrated into intention recognition. Finally, a deep-learning-based unified recognition model for driver intention is developed. This model uses a hierarchical recognition approach for speed intentions and includes a double-layer networks architecture with long short-term memory for the recognition of lateral intention. The validation results show that the created driver intention recognition model can accurately and stably recognize both speed and lateral intentions in complex driving environments.