Real-Time Anomaly and Incident Detection Based on YOLO and Lucas–Kanade Optical Flow Tracking

This paper presents a novel approach to real-time anomaly and incident detection using YOLOv26, RT-DETR, Grad-CAM, and Soft-NMS. The system aims to identify incidents, specifically accidents, by processing image datasets through advanced object detection techniques. The YOLOv26 and RT-DETR models are employed for detecting objects in images, with Soft-NMS applied to enhance the accuracy of object localization by reducing redundant predictions. Grad-CAM is utilized for visual interpretability, providing insights into the decision-making process of the models by generating class activation maps. The detection process focuses on a single class, accidents, and optimizes the performance by fine-tuning hyperparameters and leveraging the strengths of these advanced models. The methodology improves the robustness and precision of incident detection in real-time applications, offering a scalable solution for automated surveillance and safety systems.

Keywords: YOLOv26, RT-DETR, Grad-CAM, Soft-NMS, real-time anomaly detection, accident detection, object detection, hyperparameter optimization, machine learning, deep learning.