Semantic Segmentation using Modified U-Net for Autonomous Driving

Scene understanding of urban streets is a crucial component in perception task of autonomous driving application. Semantic segmentation has been extensively used in scene understanding which further provides assistance in subsequent autonomous driving tasks like object detection, path planning and motion control. But, accurate semantic segmentation is a challenging task in computer vision. U-Net is a popular semantic segmentation network used for segmentation task. In this paper, we improve the accuracy of U-Net model by replacing its encoder part with Convolution Neural Network (CNN) architecture. We compared the performance of VGG-16 and ResNet-50 CNNarchitectures. Extensive analysis was performed on Cityscapes dataset and the results demonstrated U-Net with VGG-16 encoder shows better performance than ResNet50 encoder. The model is compared with semantic segmentation CNN architectures like Fully Convolutional Network (FCN) and SegNet with mean Intersection over Union (mIoU) improved by 2%.