Real-Time traffic accident detection system on hybrid data with YOLOv9 and YOLOv11 architectures

Authors : Zeynep Balkaya, Cemil Özgültekin, Soydan Serttaş, Çiğdem Bakır

Pages : 1542-1558

Doi:10.28948/ngumuh.1715199

View : 119 | Download : 136

Publication Date : 2025-10-15

Article Type : Research Paper

Abstract :Traffic accidents have become a significant problem in our country and many other countries. Accidents increase the economic and health costs as well as the loss of life. Therefore, the timely detection of accidents is a very important issue. In this study, an accident detection and precaution system was developed on a hybrid dataset using the originally created YOLOv9 and YOLOv11 models in order to prevent these problems. In the first stage of the study, data was prepared for the training of both models by extracting similar images and labels. Hybrid datasets created with images obtained from different sources containing the \\\"accident\\\" and \\\"non-accident\\\" classes were used. In order to eliminate data imbalance, synthetic images were produced with Generative Adversarial Network (GAN), images were labeled and resized to appropriate sizes, and similar and repetitive ones were cleaned. Hierarchical Feature Attention Layer (HFAM) was added to the YOLOv9 model to better capture features, and Dynamic Context Enrichment Layer (DCEL) layer was added to the YOLOv11 model, which increases its sensitivity to environmental factors. Model performances are evaluated on five different scenarios (5 different datasets) with metrics such as mAP50, mAP50-95, accuracy, precision, sensitivity and F1 score, with hyperparameter optimization and k-fold cross validation. In this study, the performance of the proposed models is compared with classical object detection models such as SSD and Fast R-CNN. As a result, it is seen that both models can successfully detect traffic accidents and have high generalization abilities on different data structures.
Keywords : YOLO, Derin Öğrenme, mAP, Trafik Kazaları, Gerçek-Zamanlı

ORIGINAL ARTICLE URL