Gaze-based Pedestrian Safety Prediction using Attention LSTM
J Divya, M Divyadharshini, V Divyapriya · 2025
A real-time framework is proposed for predicting pedestrian intent and enhancing urban traffic safety by integrating gaze direction and motion data through an attention-based Long Short-Term Memory (LSTM) network. Unlike traditional motion-only approaches, the system incorporates visual attention cues to improve early recognition of pedestrian behavior. YOLOv5 is utilized for real-time detection of critical road elements such as vehicles, pedestrians, and zebra crossings. The combined use of gaze tracking and motion forecasting leads to significant improvements in trajectory prediction accuracy and environmental awareness. With detection and prediction accuracies of 93.5% and 92.6%, respectively, the system demonstrates robustness under varied lighting and crowd conditions. Computationally efficient and modular design enables seamless integration into intelligent transportation and urban surveillance infrastructures. Extensive evaluations demonstrate its capability to function effectively in real-world, high-traffic video surveillance environments.