Pantnagar Journal of Research

Abstract

This paper presents VRU Neural Sentinel, an adaptive neural network architecture for real time prioritization of Vulnerable Road Users (VRUs) in autonomous vehicle (AV) perception systems. The system integrates multimodal behavioral risk signals — specifically gait and movement pattern analysis with head pose and gaze direction estimation — to dynamically score and rank detected VRUs by collision risk, enabling proactive AV decision making. The proposed architecture achieves 98.8% decision accuracy and 0.00% collision rate in simulation (over 500 timesteps under mixed scenario conditions; this result should be interpreted in the context of the limited simulation scope and does not constitute a real world safety benchmark), exceeding the stated targets of 97% and 2% respectively. The system processes scenes at approximately 32ms latency (10Hz operational cycle), enabling real time response. Key innovations include a dualen coder behavioral signal pipeline, multihead crossVRU attention, vulnerability weighted risk scoring, and online adaptive threshold tuning. Simulations over 500 timesteps with mixed scenario populations (distracted pedestrians, cyclists crossing, children, elderly individuals, and wheelchair users) demonstrate consistent prioritization of highest risk individuals across all environmental conditions.