CrashSenseAI: latency-aware edge AI system for real-time traffic accident detection and notification

Abstract

Road accidents remain a critical public safety issue, where detection often depends on witnesses calling authorities, leading to delays of several minutes before police and ambulances are dispatched. This study presents CrashSenseAI (CSAI), an Edge artificial intelligence (Edge AI) accident detection and notification system built on You Only Look Once version 8 (YOLOv8) and YOLOv12 models. Using a labeled dataset of accident and non-accident images, we trained Nano, Small, and Medium variants with batch sizes of 16 and 32 to evaluate accuracy–latency trade-offs. Results show YOLOv12n achieved the lowest inference latency (3.6–3.7 ms) with high precision (0.89), while YOLOv8s provided the best balance of accuracy and efficiency (Recall 0.85, Precision 0.90, mean average precision (mAP) at an Intersection over Union threshold of 0.5 (mAP@0.5 = 0.89) and mAP@0.5–0.95 = 0.70) at 4.8–5.0 ms. Larger variants improved mAP@0.5 scores (up to 0.90) but incurred higher latency (11–14 ms). CSAI captures timestamped frames with latitude and longitude, automatically notifying police, ambulances, and nearby hospitals to speed response and scene clearance, reduce traffic congestion, and improve victim survival during the golden hour. By combining real-time detection, geolocation, evidence generation, and automated alerting, CSAI demonstrates practical scalability for AI-driven intelligent transportation and road safety applications.