Indoor Semantic Segmentation for Robot Navigating on Mobile

In recent years, there have been many successes of using Deep Convolutional Neural Networks (DCNNs) in the task of pixel-level classification (also called 'semantic image segmentation'). The advances in DCNN have led to the development of autonomous vehicles that can drive with no driver controls by using sensors like camera, LiDAR, etc. In this paper, we propose a practical method to implement autonomous indoor navigation based on semantic image segmentation using state-of-the-art performance model on mobile devices, especially Android devices. We apply a system called 'Mobile DeepLabv3', which uses atrous convolution when applying semantic image segmentation by using MobileNetV2 as a network backbone. The ADE20K dataset is used to train our models specific to indoor environments. Since this model is for robot navigating, we re-label 150 classes into 20 classes in order to easily classify obstacles and road. We evaluate the trade-offs between accuracy and computational complexity, as well as actual latency and the number of parameters of the trained models.