Walking = Traversable? : Traversability Prediction via Multiple Human Object Tracking under Occlusion
Authors:
Jonathan Tay Yu Liang,
Kanji Tanaka
Abstract:
The emerging ``Floor plan from human trails (PfH)" technique has great potential for improving indoor robot navigation by predicting the traversability of occluded floors. This study presents an innovative approach that replaces first-person-view sensors with a third-person-view monocular camera mounted on the observer robot. This approach can gather measurements from multiple humans, expanding it…
▽ More
The emerging ``Floor plan from human trails (PfH)" technique has great potential for improving indoor robot navigation by predicting the traversability of occluded floors. This study presents an innovative approach that replaces first-person-view sensors with a third-person-view monocular camera mounted on the observer robot. This approach can gather measurements from multiple humans, expanding its range of applications. The key idea is to use two types of trackers, SLAM and MOT, to monitor stationary objects and moving humans and assess their interactions. This method achieves stable predictions of traversability even in challenging visual scenarios, such as occlusions, nonlinear perspectives, depth uncertainty, and intersections involving multiple humans. Additionally, we extend map quality metrics to apply to traversability maps, facilitating future research. We validate our proposed method through fusion and comparison with established techniques.
△ Less
Submitted 29 September, 2023;
originally announced October 2023.
Active Robot Vision for Distant Object Change Detection: A Lightweight Training Simulator Inspired by Multi-Armed Bandits
Authors:
Kouki Terashima,
Kanji Tanaka,
Ryogo Yamamoto,
Jonathan Tay Yu Liang
Abstract:
In ground-view object change detection, the recently emerging mapless navigation has great potential to navigate a robot to objects distantly detected (e.g., books, cups, clothes) and acquire high-resolution object images, to identify their change states (no-change/appear/disappear). However, naively performing full journeys for every distant object requires huge sense/plan/action costs, proportio…
▽ More
In ground-view object change detection, the recently emerging mapless navigation has great potential to navigate a robot to objects distantly detected (e.g., books, cups, clothes) and acquire high-resolution object images, to identify their change states (no-change/appear/disappear). However, naively performing full journeys for every distant object requires huge sense/plan/action costs, proportional to the number of objects and the robot-to-object distance. To address this issue, we explore a new map-based active vision problem in this work: ``Which journey should the robot select next?" However, the feasibility of the active vision framework remains unclear; Since distant objects are only uncertainly recognized, it is unclear whether they can provide sufficient cues for action planning. This work presents an efficient simulator for feasibility testing, to accelerate the early-stage R&D cycles (e.g., prototy**, training, testing, and evaluation). The proposed simulator is designed to identify the degree of difficulty that a robot vision system (sensors/recognizers/planners/actuators) would face when applied to a given environment (workspace/objects). Notably, it requires only one real-world journey experience per distant object to function, making it suitable for an efficient R&D cycle. Another contribution of this work is to present a new lightweight planner inspired by the traditional multi-armed bandit problem. Specifically, we build a lightweight map-based planner on top of the mapless planner, which constitutes a hierarchical action planner. We verified the effectiveness of the proposed framework using a semantically non-trivial scenario ``sofa as bookshelf".
△ Less
Submitted 24 October, 2023; v1 submitted 26 July, 2023;
originally announced July 2023.
