Use of Deep Learning for Road Safety Education to Citizens

Massive amount of available data in the society have the capability of strengthening the electronic government. Nowadays closedcircuit television cameras of buildings, shops, and traffic lights record a huge amount of data from streets. As a novel idea, deep learning may benefit from these images to educate basic concepts of road safety to the citizens. The purpose of this study is to segment the sidewalk in street view dataset. CamVid which is a public dataset is selected to train, validate, and test the network model. An open source Python library called segmentation_models is utilized which includes four network models and ۲۵ backbones. Different network models and backbones are evaluated. For the best obtained network model and backbone, optimal values for learning rate and number of epochs parameters are investigated. U-Net network model as well as DenseNet-۲۰۱ offer better performance among other network models and backbones. Learning rate of ۰.۰۰۰۱ and ۴۰ epochs of training gave optimal segmentation accuracy equivalent to mean IoU score of ۰.۸۳۷۶۲ and mean F-score of ۰.۹۰۸۶۳. This method may be implemented on other parts of streets. Results can be offered used in games, animations, and videos.