Research On Vehicle Detection And Classification Using DNNs

Vehicle detection and classification has contributed to the development of numerous intelligent transportation systems to improve the quality of traffic service and efficiency of traffic facilities,as well as ensuring traffic safety and reducing traffic congestion.A significant number of research has been conducted in classifying and detecting vehicles because of the significance role they play in managing transport,analyzing traffic and monitoring security systems.In this study,we present an accurate system which classifies a vehicle into one of five categories:bus,car,SUV,pickup-truck and truck.We adopted two state-of-the-art object detection methods that use deep Convolutional Neural Networks(CNNs),called YOLOv2 which is a one-stage object detection method and Faster R-CNN which is a two-stage object detection method,each object detection model was trained and evaluated on the created vehicle image dataset.In this work,we have made a comprehensive and systematic study on the theories related to the existing vehicle detection and classification technology.We discuss the recent research progress in computer vision,image processing,and deep learning with a more focus on object classification and object detection.The main part of our work includes:(1).To train a deep neural network model,a vehicle image dataset was created through the use of google images.(2).Aiming at the problem of vehicle classification,a DNN state-of-the art model was adopted for both feature extraction and fine-tuning.(3).An automatic image annotation system was designed for the purpose of training both object detection methods and evaluate them on the testing image set of our dataset.(4).The performance of both object detection methods was evaluated by comparing their average mean precision achieved during different experiments.The experimental results show that a DNN fine-tuned model on our created dataset can make the classification effect to achieve the best performance on different vehicle image.Finally,the two object detection methods trained and tested on vehicle image dataset have achieved very good results.