Railroad Segmentation Using Modified UNet

AbstractOne of the complex activities in our daily life humans do is driving. We humans handle unexpected situations like reacting to weather conditions, making decisions against the rules which make us save human life who are in danger. Thus, in this situation, we are approaching a technique artificial intelligence which is the actively researched area in autonomous driving. To succeed in this field, they will be using hardware and software components that support this function. The components like LIDAR, cameras, sensors and computer vision algorithms are need to be implemented more closely which help us to achieve autonomous driving. Generally, nowadays, in many places, trains lead to autonomous driving by detecting the rail path which means there is no any interaction of humans in driving. To detect the path, we had used encoder and decoder part, where in each part, there consists of few levels, where in each level, we had used CNN layers like convolution layer, pooling layer and convolution transpose layer. We have collected the dataset manually from Web, labeled using Photoshop and masked using MATLAB. We had compared this model with existing methods where the accuracy of our model is 92%, and segmentation is very good in our case. Thus, we have concluded that modified UNet provides better accuracy compared to all other methods.KeywordsRailroad detectionUNetSegmentationCNNConvolutionPoolingTranspose convolution