The number and intensity of car accidents around the world are rising, which makes it even more important to have smart systems that can quickly and correctly process accident pictures. We need more advanced systems to quickly find accidents, label data, and record them. Manual methods take a long time, are prone to mistakes, and are not good enough for real-time apps. This paper tries to solve these problems by suggesting a system that uses advanced machine learning (ML) models to automatically process, label, and report images of car accidents. Convolutional Neural Networks (CNNs) are used for feature extraction, YOLOv8 is used for real-time accident recognition, and Transformer-based models are used for complex multi-object labeling Global Road Accidents Dataset is much better at generalizing models than the Car Crash Dataset and the Road Traffic Accidents Dataset, which are more focused on crash types and vehicle damage. This is because it has more images and a wider range of features, such as location, severity, and weather. It was tested and found that the proposed hybrid model is more accurate than the CNN-based and YOLO-only methods, which got 94.5% and 93.7% accuracy and precision, respectively, on the Global Road Accidents Dataset. The combination model also has better accuracy for annotations and fewer fake hits. Comparative research shows that combining Transformer designs with object recognition models makes it easier to understand features and make reports. Findings from this study show that the AI-driven framework can be used to automate accident investigation processes. This is a big step towards smart traffic control and emergency response systems.