Will It Last? Learning Stable Features for Long-Term Visual Localization

Abstract

Figure 1: An overview of the proposed algorithm: A classifier is used to decide whether a particular visual feature is expected to be persistent or not. Our method uses full image information as input and helps to maintain compact stable-over-time maps that can be used for life-long localization. AbstractAn increasing number of simultaneous localization and mapping (SLAM) systems are using appearance-based localization to improve the quality of pose estimates. However, with the growing time-spans and size of the areas we want to cover, appearance-based maps are often becoming too large to handle and are consisting of features that are not always reliable for localization purposes. This paper presents a method for selecting map features that are persistent over time and thus suited for long-term localization. Our methodology relies on a CNN classifier based on image patches and depth maps for recognizing which features are suitable for life-long matchability. Thus, the classifier not only considers the appearance of a feature but also takes into account its expected lifetime. As a result, our feature selection approach produces more compact maps with a high fraction of temporally-stable features compared to the current state-of-the-art, while rejecting unstable features that typically harm localization. Our approach is validated on indoor and outdoor datasets, that span over a period of several months.

Memorable Maps: A Framework for Re-Defining Places in Visual Place Recognition

“…3). More recently, a CNN able to classify input frames as stable/unstable is trained by Dymczyk et al [44] for long term visual place recognition. Similar to [43], this work also proposes that vegetation in outdoor scenes is not suitable, but does not consider outdoor dynamic objects like cars, pedestrians, animals etc.…”

Autonomous Navigation in Inclement Weather Based on a Localizing Ground Penetrating Radar

“…Particularly for visual SLAM [1], [2], an extensive body of research explicitly focuses on addressing these types of problems. The breadth of approaches involves inpainting and removal of dynamic objects [12], [13], selecting particularly persistent landmarks for map storage [14], and selecting appearance specific landmarks [15] or map segments [16]. We avoid the problems addressed in this line of work altogether by using sub-surface features that do not suffer from frequent appearance changes and occlusions.…”

A Clutter-Resistant SLAM Algorithm for Autonomous Guided Vehicles in Dynamic Industrial Environment

“…3) Calculate Mahalanobis distance D between each of these five candidates and z i , according to (32).…”

“…In [31], a dynamic deep learning SLAM is proposed for dynamic environments, in which a convolutional neural network is combined with multi-view geometry to identify dynamic objects. Dymczyk et al develop a CNN classifier to recognize stable features on the basis of their expected lifetime [32]. Demim et al point out that extraction often produces a high spatial uncertainty due to the sparse information of the raw laser LiDAR data compared to images, and they propose a new approach to filter features within a certain temporal window [33].…”

“…They also develop an adaptive SVSF-SLAM solution for dynamic environments, which can detect and remove dynamic parts of measurements [34]. In industrial environments with large size area, appearance-based maps are too large to handle [32] while sparse features are more suitable for SLAM solution, lowering computational complexity. Line features are widely employed for SLAM in many indoor scenes.…”