Abstract:
Vehicle type recognition is important for supporting traffic management and
urban planning, as Bangladesh sees rapid growth of vehicular traffic. In this
study, I provide a dataset and methodology for building a Bangladeshi vehicle
type recognition model using YOLOv9 variant. The dataset gathered from
traffic signal points in Bangladesh contains 281 images belonging to 12 vehicle
classes which has been augmented to 402 images by techniques such as image
augmentation (eg; horizontal flipping, adjusting brightness, etc). I applied
preprocessing steps like auto orientation, resize to 256x256 and histogram
equalization to improve data quality. I trained Google Colab YoloV9-C,
YoloV9-E, and YoloV9-Gelan C with batch size of 32, for 100 epochs, then
evaluated them based on mean average precision (mAP). Among all the models,
YoloV9-E could achieve the best mAP of 73.66%, which indicates that it was
able to perform well in real-time vehicle detection. Based on these insights, the
trained models were deployed on Streamlit for testing in real-world Bangladeshi
traffic environments.
