Semantic Scene Segmentation for Indoor Robot Navigation via Deep Learning

摘要

This paper presents a vision-based framework for indoor robot navigation which exploits semantic segmentation and deep learning towards accurate and efficient indoor scene mapping and collision-free navigation for hardware constrained robotics. Firstly, a scheme for accurate and efficient path extraction using deep convolutional neural networks (DCNNs) and transfer learning for semantic pixel-wise segmentation is put forward. Secondly, multiple DCNN architectures and semantic segmentation techniques are explored to highlight the challenges associated with implementation as well as the trade-offs between accuracy and efficiency associated with the state-of-the-art. Finally, the achieved models are deployed and experimentally validated. Experimental results highlight promising potential with good segmentation accuracies and real-time feasibility. Results further highlight significant accuracy-efficiency trade-offs which are strongly driven by model decoder sub-network design.