Bangladeshi Vehicle Classification using Transfer Learning with YOLOv7

Precisely detecting and classifying distinct vehicle types are essential in numerous automated systems, including traffic monitoring and vehicle detection, which are of utmost significance in computer vision. Surprisingly, deep learning methods for vehicle detection are already widely accepted. Still, their applicability to Bangladeshi vehicles has yet to be thoroughly studied, and it is crucial to develop an intelligent transportation system (ITS) in Bangladesh. Real-time vehicle detection is a crucial goal, and the You Only Look Once (YOLO) model has proven to be the most influential architecture to accomplish this goal. To determine which model is more effective in precisely classifying and detecting local vehicles in Bangladesh, we analyzed the performance of different variations of three well-known object detection YOLO-based frameworks, Scaled YOLOv4, YOLOv5 and YOLOv7. In this study, we utilized a self-produced dataset that contained frontal-view images of vehicles seen on Bangladeshi roads, including bicycles, rickshaws, motorcycles, cars, and other comparable forms of transportation. We collected videos in daylight for the dataset and then manually extracted the pictures from the videos. Our dataset included approximately 5,000 images divided into 20 categories of vehicles. The study’s results showed that YOLOv7 performs more accurately than YOLOv5 and Scaled YOLOv4 overall in terms of precision, recall, and mAP and has proven to achieve a 80.2% mAP for the classification of vehicles.