Vehicle detection and type classification in low resolution congested traffic scenes using image super resolution.

Vehicle detection and classification in real world highly congested traffic scenario is an important yet difficult task, especially because surveillance footage is captured in low resolution. Most available datasets contain high quality images and the models developed based on them do not perform well in real world scenario. In this paper, we present a Convolution Neural Network (CNN) based approach for vehicle type classification in congested vehicle traffic scenario with low resolution surveillance images. We have used highly congested low resolution CityCam dataset. To improve the classification performance of the proposed model, image super resolution technique is adopted to enhance essential details needed for differentiating classes. The experimental results demonstrate that the proposed CNN model performs better on the dataset compared with other existing state of art deep learning models. The proposed model achieved higher accuracy of 87.56% over the state of art models VGG16 and Inceptionv3 which achieves 84.62% and 84.87% respectively. Further, by enhancing the dataset with Residual Dense Network based image super resolution technique, the classification accuracy has increased around 1.73% to 89.29%. [ABSTRACT FROM AUTHOR]