Predicting Pedestrian Crossing Intentions in Adverse Weather With Self-Attention Models

The enhancement of the vehicle perception model represents a crucial undertaking in the successful integration of assisted and automated vehicle driving. By enhancing the perceptual capabilities of the model to accurately anticipate the actions of vulnerable road users, the overall driving experience can be significantly improved, ensuring higher levels of safety. Existing research efforts focusing on the prediction of pedestrians' crossing intentions have predominantly relied on vision-based deep learning models. However, these models continue to exhibit shortcomings in terms of robustness when faced with adverse weather conditions and domain adaptation challenges. Furthermore, little attention has been given to evaluating the real-time performance of these models. To address these aforementioned limitations, this study introduces an innovative framework for pedestrian crossing intention prediction. The framework incorporates an image enhancement pipeline, which enables the detection and rectification of various defects that may arise during unfavorable weather conditions. Subsequently, a transformer-based network, featuring a self-attention mechanism, is employed to predict the crossing intentions of target pedestrians. This augmentation enhances the model's resilience and accuracy in classification tasks. Through evaluation on the Joint Attention in Autonomous Driving (JAAD) dataset, our framework attains state-of-the-art performance while maintaining a notably low inference time. Moreover, a deployment environment is established to assess the real-time performance of the model. The results of this evaluation demonstrate that our approach exhibits the shortest model inference time and the lowest end-to-end prediction time, accounting for the processing duration of the selected inputs.