Bird’s-Eye-View (BEV) segmentation is a essential technology for safe and efficient navigation. This is even more necessary in indoor driving, where there are dynamic and unstructured objects such as people and robot. However, since there is no way to generate training data, most of the researches have been conducted mainly in outdoor environments. In this paper, we propose an innovative approach to address this challenge. We automatically generate BEV training data for indoor environments based on the physics engine of a simulator. This eliminates the needs for tons of real data and virtual environments. We also propose a lightweight network architecture capable of BEV panoptic segmentation in real-time. Based on this network and simple image processing, our framework shows fast but also robust performance in real-world environments with no gap between simulation and reality. The network can inference at a speed of about 61 FPS on AGX Orin in FP16 mode. Furthermore, it outperforms existing algorithms in the semantic segmentation task, achieving 14% higher mIoU.

This paper presents a new coarse-to-fine global localization approach based on place learning with a 2-D range scan; we use a support vector machine (SVM) to train a set of classifiers for place recognition. Exploiting the graphical property of generalized Voronoi graph extracted from an occupancy grid map, we segment the map into local places with a spectral clustering algorithm and wavefront propagation algorithm. A bag-of-words approach is employed to build feature vectors for training the SVM classifiers, which utilizes the shape context descriptors computed from each scan point. The global localization process consists of two stages: coarse localization and fine localization. Coarse localization from the SVM-based place recognition selects candidate places where the robot may be located, and fine localization computes the relative poses to the candidate places with fast spectral scan matching and estimates the correct robot pose with a particle filter algorithm, in which the candidate poses are exploited as seeds for initial random sampling. The feasibility of the proposed methods is demonstrated by the results of extensive simulations and experiments.

This paper presents a fast and efficient computational approach to higher order spectral graph matching. Exploiting the redundancy in a tensor representing the affinity between feature points, we approximate the affinity tensor with the linear combination of Kronecker products between bases and index tensors. The bases and index tensors are highly compressed representations of the approximated affinity tensor, requiring much smaller memory than in previous methods, which store the full affinity tensor. We compute the principal eigenvector of the approximated affinity tensor using the small bases and index tensors without explicitly storing the approximated tensor. To compensate for the loss of matching accuracy by the approximation, we also adopt and incorporate a marginalization scheme that maps a higher order tensor to matrix as well as a one-to-one mapping constraint into the eigenvector computation process. The experimental results show that the proposed method is faster and requires smaller memory than the existing methods with little or no loss of accuracy.

This paper presents a new scan matching method for mobile robot localization. The proposed method is based on a spectral technique, which finds consistent correspondences between two range scans without any initial alignment while considering how well their associated pairwise geometric relationships are satisfied. Based on the scan matching method, we also suggest a strategy for global localization in indoor environments, which is applicable to both grid maps and topological maps having metric information. The feasibility of the proposed methods is demonstrated by experimental results.

This paper proposes a novel global localization approach that uses hybrid maps of objects and spatial layouts. We model indoor environments using the following visual cues from a stereo camera: local invariant features for object recognition and their 3D positions for object location representation. We also use a 2D laser range finder. Therefore, we can build a hybrid local node for a topological map that is composed of an object location map and a spatial layout map. Based on this modeling, we suggest a coarse-to-fine strategy for the global localization. The coarse pose is obtained by means of object recognition and point cloud fitting, and then its fine pose is estimated with a probabilistic scan matching algorithm. With real experiments, we show that our proposed method can be an effective global localization algorithm.