Performance Analysis of Convolution Neural Networks Using Semantic Segmentation for Driving Scenes

Since past decade many real time system have been developed. Technologies such as autonomous vehicles, virtual reality systems, drones have been using application of machine learning techniques to perform there task. Thus involves observation, planning, as well as execution in ever-changing situation, safety and accuracy are important factors. The focus of this study is to segment images by utilizing deep learning techniques in order to aid in better understanding of driving scene perception in order to help autonomous driving systems in distinguishing between ground reality and prediction. Deep learning models such U-Net, FCN and FPN are best among state-of-art method that are used for providing solutions to real world problems. For this research U-Net architecture and FPN architecture were used on camvid driving scene dataset and image segmentation was performed to understand the driving scenes to analyse the difference in ground reality and machine result. Both models showed good results in comparison to other methods. U-Net model is applied with ResNet-50 and FPN with ResNeXt networks for segmentation on driving scene, where both models were evaluated on basis of IoU and Dice loss. U-Net achieved IoU score of 82% and FPN architecture model achieved IoU score of 84%