Scene Analysis For Indoor Robot Navigation

Abstract

Making mobile robots truly ubiquitous and cohabit with human beings require enhancing robot’s ability to understand complex indoor environments. Service robots must perceive and understand complex indoor scenes and be able to recover room layout to better grasp the space and orientation of objects and 3D surfaces in the room. Robots ability to reason about the 3D surface have great implication for navigation and object detection. Robots must also take advantage of widely available information in indoor environment (i.e. text, sign door etc.) to solve long standing navigation problems like Loop closure problem (i.e. robot’s inability to recognize a place it already visited. There has been very little work directed toward employing scene analysis algorithm for robot navigation and using text and sign to solve loop closure problem. So, introducing mechanism that integrate scene analysis and object detection algorithm will solve loop closure problem and improve performance of mobile robot navigation. This research proposes integrating scene analysis algorithm called room layout recovery and Aggregate Channels Features (ACF) object detector for mobile robot navigation. The algorithm first recover, classify, segment and label geometric surfaces (walls, ceiling and floors). Then the output of the algorithm is used to train Aggregate Channels Features (ACF) object detector. Which will be used to detect text, sign and doors. The algorithm is then implemented on turtle bot and simulated world to evaluate the effectiveness of the proposed algorithm. The proposed algorithm of integrating scene analysis and Aggregate Channels Features (ACF) achieves average precision of 0.7 & log average miss rate of 0.4 when door is partially visible and average precision of 1.0 and log average miss rate of 0.0 when door is fully visible. Additionally, the proposed algorithm achieves average precision of 0.9 and log average miss rate of 0.2 for text and sign detection. Moreover, the algorithm is tested using turtle bot in simulated world where it successfully detects door, text and sign. This research shows the importance of integrating scene analysis and object detection for robot navigation.