Multi-Robot Simultaneous Localization and Mapping with Generalized Objects – Tackling Unstable Communication and Tracking Maneuver Switching Objects

為 了 實 現 動 態 環 境 中 的 機 器 人 同 時 定 位 與 建 地 圖 （ Simultaneous Localization and Mapping 或SLAM），研究學者提出了同時定位、建地圖與移 動物體追蹤（Simultaneous Localization, Mapping, and Moving Object Tracking 或 SLAMMOT）的理論架構以及二種實作演算法：伴隨移動物體偵測與追蹤的同時定位與建地圖演算法、以及使用一般性物體的同時定位與建地圖演算法。在伴隨移動物體偵測與追蹤的同時定位與建地圖演算法中，對於動態與靜態物體的感測資訊會被分離，而對於動態物體的感測資訊則不會直接地被同時定位與建地圖演算法使用。相對的，在使用一般性物體的同時定位與建地圖演算法中，機器人、靜態物體以及動態物體的聯合狀態機率分布會被估算，因此同時定位與建地圖演算法將可與移動物體追蹤演算法相互幫忙。 本博士論文的主要研究目標在於將使用一般性物體的同時定位與建地圖演算法擴展至多機器人情境中，同時透過實際應用情境展示整合移動物體對於整體估測的有效性。首先，我們提出多機器人同時定位與追蹤演算法（Multi-Robot Simultaneous Localization and Tracking 或MR-SLAT）並在機器人足球賽與交通情境中驗證，我們將會展示，藉由整合移動物體，多機器人同時定位與追蹤比單一機器人自我定位以及多機器人協同定位都有更好的表現，並且，整合移動物體追蹤是解決許多定位挑戰的關鍵。更進一步地，我們將處理多機器人同時定位與追蹤演算法於通訊不穩定情境中的議題，我們提出通訊適應式多機器人同時定位與追蹤演算法（ Communication Adaptive MR-SLAT 或ComAd MR-SLAT）以結合量測分享與估測分享二種現有資訊分享方式的長處，通訊適應式多機器人同時定位與追蹤演算法會基於當前通訊狀態，透過線上動態調整資訊分享策略來最佳化演算法的表現。 此外，我們也在更具挑戰的單眼相機情境中展示整合移動物體的效果，我們將會指出，整合運動模態不確定的移動物體可能為演算法帶來整體估測準度下降的風險。對此，我們提出運動模態改變偵測演算法（Maneuver Switching Detection 或MSD）以將使用一般性物體的同時定位與建地圖演算法實現於物體運動模態可能改變的情境，其中，我們整合了追蹤領域的多運動模型方法（Multiple Model 或MM）以及吉布斯取樣法（Gibbs Sampling）以處理演算法複雜度過高的問題。在本博士論文之中，我們透過許多實驗展示整合移動物體的有效性以及我們所提出的演算法於真實應用情境中的可行性。To facilitate simultaneous localization and mapping (SLAM) in dynamic environments, the theoretical framework of simultaneous localization, mapping and moving object tracking (SLAMMOT) has been proposed, in which two solutions exist: SLAM with detection and tracking of moving objects (SLAM with DATMO) and SLAM with generalized objects (SLAM with GO). In SLAM with DATMO, measurements of moving objects are separated from those of static objects and are not utilized in the SLAM estimator. In SLAM with GO, the states of the robot, static objects and moving objects are jointly estimated, and thus SLAM and moving object tracking (MOT) can contribute to each other. The main goal of this thesis is to extend SLAM with GO to multi-robot scenarios and demonstrate the effectiveness of incorporating moving objects in practical domains. The multi-robot simultaneous localization and tracking (MR-SLAT) algorithm is proposed and evaluated in two scenarios, RoboCup and traffic. It will be demonstrated that by incorporating moving objects, MR-SLAT outperforms single robot localization and cooperative localization, and incorporating moving objects serves as the key to several localization challenges. Furthermore, for tackling practical unstable communication issues, a hybrid algorithm, communication adaptive MR-SLAT (ComAd MR-SLAT), is proposed to combine the advantages of two existing information sharing strategies, measurement-sharing and belief-sharing. In ComAd MR-SLAT, the sharing strategy is adaptively switched online according to the communication condition in order to optimize the estimation performance. The effectiveness of incorporating moving objects is also demonstrated in the challenging monocular scenario. It will be pointed out that incorporating moving objects with uncertain maneuvers suffers from the risk to downgrade the overall SLAM with GO performance. Accordingly, a maneuver switching detection (MSD) algorithm is proposed to extend monocular SLAM with GO to practical scenes in which the maneuvers of objects can switch. To integrate the multiple model (MM) approach from the tracking literature, a Gibbs sampling based MSD algorithm is designed to tackle the exponential complexity issue. Through ample experiments in this thesis, the practical applicability of the proposed algorithms and the effectiveness of incorporating moving objects are demonstrated