Schatting en tracking van de 6D-pose van objecten voor robotwaarneming

Estimating the 6D pose of objects, comprising both 3D position and orientation, is a key requirement for many robotic applications. It enables robots to perceive and interact intelligently with the physical environment, supporting tasks in human-robot collaboration, logistics automation, and industrial operations. In human-robot collaborative assembly settings, accurate and continuous 6D pose tracking of product parts is essential for precise grasping and placement. Moreover, tracking the trajectories of parts and tools provides valuable contextual information for monitoring task progress and inferring operator intent. In logistics warehouses, effective object handling requires robots to estimate and track the 6D pose of items such as packages and boxes, which are often stacked irregularly or placed at varying orientations. As a core perceptual capability, 6D pose estimation and tracking bridge raw sensory input with high-level robotic decision-making, enabling robustness, precision, and adaptability in complex real-world settings. This thesis addresses diverse challenges in 6D object pose estimation and tracking for robotic perception, aiming to develop methods that are robust, adaptable to diverse input constraints, and deployable across a wide range of robotic applications. In many industrial applications, the availability of computer-aided design (CAD) models for target objects enables instance-level 6D object pose tracking. The first contribution of this thesis is a comprehensive survey of techniques in this domain. Unlike single-image-based 6D pose estimation methods that process each frame independently, instance-level 6D pose tracking methods exploit temporal coherence across frames to improve robustness, accuracy, and computational efficiency. Despite its significance, there is a noticeable lack of survey studies covering recent advancements in this area. To address this gap, we present a thorough review of the latest developments in instance-level 6D object pose tracking. The survey provides a systematic examination of metrics, datasets, and methodologies employed in this field. Based on the problem modeling approach, existing methods are categorized into optimization-based, learning-based, filtering-based, and hybrid approaches that combine various techniques. Additionally, quantitative results on several publicly available datasets are presented and analyzed, along with insightful observations and suggestions for future research directions. The second contribution is an improved approach to instance-level 6D pose tracking in fast motion scenarios. While end-to-end deep learning-based 6D pose tracking methods have demonstrated strong performance on standard benchmarks featuring slowly varying object motions, they often struggle in fast motion scenarios. To address this limitation, we propose a novel framework that integrates a Relative Pose Estimation Network into an Extended Kalman Filter (EKF). The EKF employs a constant velocity motion model, where the measurement updates are derived from the network's output. This hybrid approach improves tracking robustness during fast movements. Evaluations on challenging hand-object interaction sequences from the Laval dataset demonstrate a significant boost in tracking accuracy, while maintaining real-time performance at over 30 frames per second. The third contribution addresses unknown object tracking and reconstruction in unstructured environments where CAD models and object categories are unavailable. In such scenarios, robots must segment, track, and reconstruct previously unseen objects in real time. To this end, we propose SegFusion, a dynamic RGB-D simultaneous localization and mapping (SLAM) system that simultaneously performs camera tracking, static scene mapping, and dense reconstruction of semantically unknown, moving objects. SegFusion combines sparse and dense visual cues within a pose graph optimization framework for low-drift tracking and employs surfel-based mapping for efficient, dense reconstruction. For segmentation, we introduce a hybrid approach that uses motion inconsistencies to generate prompts for the Segment Anything Model (SAM), enabling accurate, category-agnostic object segmentation. Evaluations on three dynamic SLAM benchmarks show that SegFusion achieves superior tracking accuracy, segmentation quality, and reconstruction quality. The fourth contribution is a practical and generalizable 6D pose estimation pipeline for industrial pallet bins, targeting autonomous forklift and mobile manipulation applications in logistics environments. The proposed modular perception paradigm combines synthetic-data-driven 2D object detection with 6D pose refinement using a transformer-based foundation model, enabling scalable pose estimation without object-specific retraining when CAD models are available at inference time. Detectors fine-tuned solely on synthetic data provide sufficiently accurate initializations for reliable pose estimation, achieving performance comparable to ground-truth mask initialization on a custom pallet-bin dataset. Experimental analysis shows robust performance across typical operating distances, with degradation occurring primarily under severe occlusion or truncation.