A Novel Distribution for Representation of 6D Pose Uncertainty

The 6D Pose estimation is a crux in many applications, such as visual perception, autonomous navigation, and spacecraft motion. For robotic grasping, the cluttered and self-occlusion scenarios bring new challenges to the this field. Currently, society uses CNNs to solve this problem. The CNN models will suffer high uncertainty caused by the environmental factors and the object itself. These models usually maintain a Gaussian distribution, which is not suitable for the underlying manifold structure of the pose. Many works decouple rotation from the translation and quantify rotational uncertainty. Only a few works pay attention to the uncertainty of the 6D pose. This work proposes a distribution that can capture the uncertainty of the 6D pose parameterized by the dual quaternions, meanwhile, the proposed distribution takes the periodic nature of the underlying structure into account. The presented results include the normalization constant computation and parameter estimation techniques of the distribution. This work shows the benefits of the proposed distribution, which provides a more realistic explanation for the uncertainty in the 6D pose and eliminates the drawback inherited from the planar rigid motion.

Robustness Improvement of Visual Templates Matching Based on Frequency-Tuned Model in RatSLAM

This paper describes an improved brain-inspired simultaneous localization and mapping (RatSLAM) that extracts visual features from saliency maps using a frequency-tuned (FT) model. In the traditional RatSLAM algorithm, the visual template feature is organized as a one-dimensional vector whose values only depend on pixel intensity; therefore, this feature is susceptible to changes in illumination intensity. In contrast to this approach, which directly generates visual templates from raw RGB images, we propose an FT model that converts RGB images into saliency maps to obtain visual templates. The visual templates extracted from the saliency maps contain more of the feature information contained within the original images. Our experimental results demonstrate that the accuracy of loop closure detection was improved, as measured by the number of loop closures detected by our method compared with the traditional RatSLAM system. We additionally verified that the proposed FT model-based visual templates improve the robustness of familiar visual scene identification by RatSLAM.

Deep Global Features for Point Cloud Alignment

Point cloud registration is a key problem in computer vision applications and involves finding a rigid transform from a point cloud into another such that they align together. The iterative closest point (ICP) method is a simple and effective solution that converges to a local optimum. However, despite the fact that point cloud registration or alignment is addressed in learning-based methods, such as PointNetLK, they do not offer good generalizability for point clouds. In this stud, we proposed a learning-based approach that addressed existing problems, such as finding local optima for ICP and achieving minimum generalizability. The proposed model consisted of three main parts: an encoding network, an auxiliary module that weighed the contribution of each input point cloud, and feature alignment to achieve the final transform. The proposed architecture offered greater generalization among the categories. Experiments were performed on ModelNet40 with different configurations and the results indicated that the proposed approach significantly outperformed the state-of-the-art point cloud alignment methods.

Registration with a small number of sparse measurements

This work introduces a method for performing robust registration given the geometric model of an object and a small number (less than 20) of sparse point and surface normal measurements of the object’s surface. Such a method is of critical importance in applications such as probing-based surgical registration, contact-based localization, manipulating objects devoid of visual features, etc. Our approach for sparse point and normal registration (SPNR) is iterative in nature. In each iteration, the current best pose estimate is perturbed to generate several candidate poses. Among the generated poses, one pose is selected as the best, by evaluating an inexpensive cost function. This pose is used as the initial condition to estimate the locally optimum registration. This process is repeated until the registration estimate converges within a tolerance bound. Two variants are developed: deterministic (dSPNR) and probabilistic (pSPNR). The dSPNR is faster than pSPNR in converging to the local optimum, but the pSPNR requires fewer parameters to be tuned. The pSPNR also provides pose-uncertainty information in addition to the registration estimate. Both approaches were evaluated in simulation using various standard datasets and then compared with results obtained using state-of-the-art methods. Upon comparison with other methods, both dSPNR and pSPNR were found to be robust to initial pose errors as well as noise in measurements. The effectiveness of the approaches are also demonstrated with robot experiments for the application of probing-based registration.