Proactive Pedestrian Risk Prediction under Dynamic Environmental Conditions

Abstract

Automated systems built to protect pedestrians on the road today depend almost entirely on flat, 2D object detection setups. The trouble is, these baseline models easily break down when forced into unpredictable, real-world driving scenarios especially during dark nighttime hours, harsh weather, or when the vehicle bounces over rough roads and shakes the camera. To bridge this gap, this paper introduces a unified, multi-stage computer vision pipeline de-signed to actively predict and avoid pedestrian risks before they happen. Our framework tack-les visual distortions upfront by pairing Contrast Limited Adaptive Histogram Equalization (CLAHE) with a real-time digital image stabilization routine. Once the video feed is stabilized, we pass the frames to a fast YOLOv8 deep learning network to pinpoint exactly where pedestrians are. Simultaneously, we feed the data into a Monocular Depth Estimation (MDE) pipeline running the MiDaS model, allowing us to pull accurate 3D depth vectors (Z-axis) straight out of a single, standard camera feed without needing expensive sensors. From there, a predictive Kinematic Risk Matrix analyses the pedestrian's frame-by-frame walking direction and speed relative to the moving car to calculate real-time Time-to-Collision (TTC) thresholds on the fly. Testing shows that our integrated system easily maintains a high-throughput processing speed of more than 30 Frames Per Second (FPS). More importantly, it successfully calculates spatial hazards 2 to 3 seconds before a potential crash would occur, proving it to be a highly dependable, practical computing architecture for modern autonomous vehicle safety.