Optimizing Appearance-Based Localization with Catadioptric Cameras: Small-Footprint Models for Real-Time Inference on Edge Devices

This paper considers the task of appearance-based localization: visual place recognition from omnidirectional images obtained from catadioptric cameras. The focus is on designing an efficient neural network architecture that accurately and reliably recognizes indoor scenes on distorted images from a catadioptric camera, even in self-similar environments with few discernible features. As the target application is the global localization of a low-cost service mobile robot, the proposed solutions are optimized toward being small-footprint models that provide real-time inference on edge devices, such as Nvidia Jetson. We compare several design choices for the neural network-based architecture of the localization system and then demonstrate that the best results are achieved with embeddings (global descriptors) yielded by exploiting transfer learning and fine tuning on a limited number of catadioptric images. We test our solutions on two small-scale datasets collected using different catadioptric cameras in the same office building. Next, we compare the performance of our system to state-of-the-art visual place recognition systems on the publicly available COLD Freiburg and Saarbrücken datasets that contain images collected under different lighting conditions. Our system compares favourably to the competitors both in terms of the accuracy of place recognition and the inference time, providing a cost- and energy-efficient means of appearance-based localization for an indoor service robot.

Real-Time Detection of Non-Stationary Objects Using Intensity Data in Automotive LiDAR SLAM

This article aims at demonstrating the feasibility of modern deep learning techniques for the real-time detection of non-stationary objects in point clouds obtained from 3-D light detecting and ranging (LiDAR) sensors. The motion segmentation task is considered in the application context of automotive Simultaneous Localization and Mapping (SLAM), where we often need to distinguish between the static parts of the environment with respect to which we localize the vehicle, and non-stationary objects that should not be included in the map for localization. Non-stationary objects do not provide repeatable readouts, because they can be in motion, like vehicles and pedestrians, or because they do not have a rigid, stable surface, like trees and lawns. The proposed approach exploits images synthesized from the received intensity data yielded by the modern LiDARs along with the usual range measurements. We demonstrate that non-stationary objects can be detected using neural network models trained with 2-D grayscale images in the supervised or unsupervised training process. This concept makes it possible to alleviate the lack of large datasets of 3-D laser scans with point-wise annotations for non-stationary objects. The point clouds are filtered using the corresponding intensity images with labeled pixels. Finally, we demonstrate that the detection of non-stationary objects using our approach improves the localization results and map consistency in a laser-based SLAM system.

Leveraging Visual Place Recognition to Improve Indoor Positioning with Limited Availability of WiFi Scans

WiFi-based fingerprinting is promising for practical indoor localization with smartphones because this technique provides absolute estimates of the current position, while the WiFi infrastructure is ubiquitous in the majority of indoor environments. However, the application of WiFi fingerprinting for positioning requires pre-surveyed signal maps and is getting more restricted in the recent generation of smartphones due to changes in security policies. Therefore, we sought new sources of information that can be fused into the existing indoor positioning framework, helping users to pinpoint their position, even with a relatively low-quality, sparse WiFi signal map. In this paper, we demonstrate that such information can be derived from the recognition of camera images. We present a way of transforming qualitative information of image similarity into quantitative constraints that are then fused into the graph-based optimization framework for positioning together with typical pedestrian dead reckoning (PDR) and WiFi fingerprinting constraints. Performance of the improved indoor positioning system is evaluated on different user trajectories logged inside an office building at our University campus. The results demonstrate that introducing additional sensing modality into the positioning system makes it possible to increase accuracy and simultaneously reduce the dependence on the quality of the pre-surveyed WiFi map and the WiFi measurements at run-time.