Enhanced Vehicle Detection and Traffic Density Analysis Using Fine-Tuned YOLOv8

The deployment of technology in road management systems is rising rapidly, enabling real-time visual information available at thousands of locations throughout road networks. In traffic accident prevention and early identification, a possible measure is the proper identification of vehicles on the road. CNN has dramatically enhanced object detection, especially when boosting the classical computer vision approaches but existing approaches have limitations. In particular, the detection rates are low for pre-trained models, especially in the case of minor objects, while full-scale networks of IP cameras require labeling of vehicles, which is virtually impossible. However, these challenges can be solved by using the approach presented in this paper, which is based on the YOLOv8 model. Starting from this work, the choice and evaluation of a pre-trained YOLOv8 model on the COCO dataset are made with special attention paid to its initial detection of vehicles. Another vehicle-specific set is then collected based on the specific vehicle the model is to detect and annotated to fine-tune the model detection on different types of vehicles. By fine-tuning from the YOLOv8 model, special consideration is given the precision and the recall in vehicle detection from aerial view angles. Many learning curves and confusion matrices along with other performance indicators are applied to test the model’s performance and its ability to generalize on new data. For achieving the inference on the validation images and the other unseen test data, this model proves to be practical for real-time traffic monitoring and traffic management.