Real-Time Traffic Vehicle Detection in Bangladesh Using Yolo

Abstract

In this thesis, the implementation of machine learningbased object detection algorithms in computer vision is discussed. The key focus is the detection and identification of different types of vehicles found in Bangladesh in realtime. Vehicle detection is one of the fundamental requirements for applications like traffic surveillance and autonomous cars. It is more challenging in Bangladesh because of the irregular traffic, numerous types of vehicles, and the lack of a healthy dataset. Different neural ne tworks were trained using tiny, and YOLO YOLOv3, YOLOv4 v5 on the “Dhaka Traffic Detection Challenge Dataset” for the classification. Both YOLO v3 and v4tiny were run on the Darknet framework, trained on a mo powered computer. YOLO v5 ran on PyTorch derately and the model was trained on Google Colaboratory, a cloudbased platform. Codes were written in Python 3.6. Roboflow, an online based computer vision application, was used to organize the dataset for training. For real-- time detection, an IP camera was used . The camera is capable of streaming video wirelessly without significant lag. The results of different models are compared. YOLOv5 performed the best among the models and produced the most promising results. The model took not more than 50 milliseconds to process each frame of the video feed on our moderately powered workstation. Which is good enough to detect vehicles in realtime. It was also able to detect more vehicles accurately that the other two models. This research has great potential and can be c onsidered a step forward towards smart traffic systems and autonomous vehicles in Bangladesh.