Towards life-long visual localization using an efficient matching of binary sequences from images

Abstract

Life-long visual localization is one of the most challenging topics in robotics over the last few years. The difficulty of this task is in the strong appearance changes that a place suffers due to dynamic elements, illumination, weather or seasons. In this paper, we propose a novel method (ABLE-M) to cope with the main problems of carrying out a robust visual topological localization along time. The novelty of our approach resides in the description of sequences of monocular images as binary codes, which are extracted from a global LDB descriptor and efficiently matched using FLANN for fast nearest neighbor search. Besides, an illumination invariant technique is applied. The usage of the proposed binary description and matching method provides a reduction of memory and computational costs, which is necessary for long-term performance. Our proposal is evaluated in different life-long navigation scenarios, where ABLE-M outperforms some of the main state-of-the-art algorithms, such as WI-SURF, BRIEF-Gist, FAB-MAP or SeqSLAM. Tests are presented for four public datasets where a same route is traversed at different times of day or night, along the months or across all four seasons.

Robust Multimodal Sequence-Based Loop Closure Detection via Structured Sparsity

“…Deep features based on convolutional neural networks (CNNs) were adopted to match image sequences [31]. Global features can encode whole image information and no dictionary-based quantization is required, which showed promising performance for long-term place recognition [2,26,27,30,33].…”

“…For example, the Chow Liu tree was used by the FAB-MAP SLAM [6,7]. The KD tree was implemented using FLANN to perform fast nearest neighbor search in the RTAB-MAP [21,22] and some other methods [2,21] for efficient image-to-image matching. Very recently, methods based on sparsity-inducing norms were introduced to decide the globally most similar template to the query image [24] (details in Section III-A).…”

“…For example, GIST features [24], built from responses of steerable filters at different orientations and scales, were applied to perform place recognition [37]. The Local Difference Binary (LDB) features were used to represent scenes by directly computing a binary string using simple intensity and gradient differences of image grid cells [2]. The SeqSLAM approach [26] utilized the sum of absolute differences between contrast low-resolution images as global features to perform sequence-based place recognition.…”

“…Most previous techniques applied global [2,24,26] or local [1,7,22,29] visual features to represent new observations and the scene templates of previously visited places; then a newly observed scene is matched with a scene template using a similarity score or a nearest neighbor search [2,6]. However, loop closure detection problems become very challenging for the methods to solve robustly when two places look similar, which is often referred to as perceptual aliasing [14], or when the same location exhibits significant appearance changes during a long-term SLAM operation (e.g., in different illumination, weather, and vegetation conditions across months or seasons [18,26]), often called long-term place recognition.…”

“…Another promising direction of research to address the challenges is based on the idea of integrating temporal information and consistency by matching sequences of templates and observed frames, instead of individual images. Sequencebased loop closure detection methods demonstrated a high robustness to perceptual aliasing as well as substantial perceptual changes due to weather, daylight, and season changes in longterm loop closure detection [2,17,18,26].…”