Abstract
Road traffic accidents in Dhaka are among the worst in the world, along with huge human fatalities and vast economic losses. An advanced car accident detecting system using a YOLOv11 model is proposed for highly efficient real-time detection to further upgrade the emergency response systems. Moreover, a comparative analysis has been made on YOLOv9, YOLOv10 and YOLOv11 on the same dataset to detect accident accurately. The pre-trained weights were used through state-of-the-art object detection techniques such as IoU and NMS, which are very accurate for the YOLOv11. The dataset of 9000 labelled images have been used for training for extremely accurate object detection and classification of on-road accidents. On a 0.8249 Recall, it does an exceptionally good job with a perfect Precision of 1.0000 and a mean Average Precision of 0.9940 on a 50 % IoU threshold. It also has very low latency, requiring only 19.93 ms per frame on a GPU, hence suitable for real-time applications. This may potentially improve the response time at an accident site, significantly reducing the risk of fatalities. These results point toward the great avenue that AI-driven systems can offer in the quest toward making roads safer and responses better. The integration of such traffic management within existing frameworks would ensure great monitoring and timely response actions to save a life. It also reduces the impact on society resulting from traffic accidents. Indeed, this is a giant step towards harnessing artificial intelligence potentials for public safety and infrastructural development.