This study introduces a novel feature-based long short-term memory (LSTM) neural network designed to predict the severity of pedestrian–vehicle interactions at unsignalized intersections, aiming to enhance traffic safety significantly. The LSTM model dynamically integrates real-time data variables such as pedestrian and vehicle speeds, distances, and types, enabling it to accurately predict potential hazardous interactions in mixed traffic conditions. Utilizing a comprehensive data set of over 4,300 pedestrian–vehicle interactions, the interactions were meticulously categorized from “Safe Passage” to “Conflict” based on detailed behavioral analyses. Training and validation of the LSTM model demonstrated high accuracy and strong generalization capabilities. The model demonstrated strong predictive capabilities across varying levels of pedestrian–vehicle interaction severity. It performed well in identifying Safe Passage interactions, with high precision and a balanced trade-off between precision and recall for Critical Events. The model also excelled in predicting conflicts, showing robustness in distinguishing these critical interactions. Further evaluation showed significant improvement over baseline models and consistent performance across different urban environments. After iterative retraining with external data, the model’s overall performance improved significantly, reflecting its capacity to generalize about diverse traffic conditions. The successful deployment of this LSTM model marks a significant advancement in real-time traffic safety technology. Its ability to provide accurate predictions of interaction severity offers a proactive tool for providing real-time warnings to mitigate pedestrian–vehicle collisions, highlighting its potential for integration into traffic management and autonomous vehicle systems. This study underscores the model’s applicability across various traffic environments, making it a vital asset for urban safety enhancements.
Muduli et al. (Wed,) studied this question.