HO3-SLAM: Human-Object Occlusion Ordering as Add-on for Enhancing Traversability Prediction in Dynamic SLAM

Detecting traversable regions from monocular vision is a fundamental problem for indoor robot navigation. In existing works, this is typically formulated as the problem of reconstructing floor regions that are directly visible to the robot. However, the assumption of floor visibility is often violated in crowded scenes such as office environments, which make the existing solutions less effective. To complement existing solutions, we propose a novel approach based on pedestrian observation, i.e., observing the behaviour of pedestrians assuming that the floor region they are standing on is traversable. In our observation, it is recently becoming realistic to implement this new approach by combining state-of-the-art of computer vision techniques, including (1) detection of dynamic objects (e.g., YOLO object detector), (2) absolute localization of stationary objects (e.g., SLAM), and (3) relative localization between dynamic and stationary objects (e.g., Occlusion ordering). A prototype system was implemented by extending state-of-the-art techniques of dynamic SLAM and object detector models, and its basic performance was demonstrated through real-world experiments. We believe that our approach can offer a valuable solution for challenging application domains such as human robot collaboration and long-horizon action planning.

I. Introduction

Detecting walkable areas using monocular cameras is crucial for safe and efficient robot navigation. This problem has been extensively studied in computer vision, with various approaches such as obstacle detection [1], free space detection [2], and 3D reconstruction [3]–[5]. Many researchers have used deep learning models to predict pixel-level depth values from images [1], while others have developed vision-based algorithms that gradually expand into free space from the robot's feet [2]. Additionally, some studies have focused on reconstructing 3D geometric models using the robot's onboard camera view sequence [3]–[5]. However, most of the existing studies assume that the floor area is directly visible, which is not always the case in crowded scenarios as shown in Fig. 1. For instance, in crowded scenes like an office where humans and robots collaborate [6], obstacles (e.g., desks) often occlude the floor areas humans are standing on. In such situations, floor detection methods become ineffective, and neither obstacle detection nor 3D reconstruction methods can provide floor information.