Visual positioning is one of the important research directions of indoor positioning algorithms and systems. Image feature descriptor plays a key role in most visual positioning algorithms, which directly affects the speed and accuracy of positioning. This paper focuses on the application of image feature descriptors in visual positioning and studies image feature detection and extraction algorithms. The image feature descriptors are divided into three categories according to the characteristics of different, namely local gradient-based descriptors, image intensity-based descriptors, and learning-based descriptors. Which steps, characteristics, applicable scene, and application in visual positioning or image matching of the methods are studied. The main purpose of this paper is to make a review of the image feature descriptors that may be used in visual positioning, and provide a reference for the research and innovation of visual indoor positioning using image feature descriptors. Image features; feature descriptors, visual positioning Proceedings International Conference on Indoor Positioning and Indoor Navigation, 29 Nov. - 2 Dec. 2021, Lloret de Mar, Spain
The demand for location information is gradually increasing in daily life [
In the research of indoor visual positioning, the method using image features has achieved good
results. The location, shape, and color of some objects change less indoors, which provides a good
foundation for indoor positioning based on image features [
2021 Copyright for this paper by its authors.
Image feature descriptors have been widely studied and applied in many fields, such as image
matching, image retrieval, image classification, object recognition, target tracking, change detection,
and so on [
Compared with other reviews such as [
The following part of this paper summarizes the extraction methods and various descriptors according to the classification of image feature descriptors. The application of various descriptors will be summarized and the indoor positioning method based on image feature descriptors outlook will be discussed.
The generation process of image feature descriptors includes two stages: the extraction of image
feature points and the description of image feature points [
The image feature point extraction algorithm is the method to find robust and stable feature points
in the image [
Harris et al. proposed the Harris corner detection algorithm in 1988 [
Feature description refers to the process of generating descriptions based on the extracted feature points according to the pixel gray level or image texture around the feature points. The feature descriptors are often stored in the form of vectors or matrices. We will focus on the influence of different description methods of image feature points on the matching performance of image feature descriptors in indoor visual positioning. According to the calculation method, feature descriptors are divided into three types: local gradient-based descriptors, image intensity-based descriptors, and learning-based descriptors.
Image feature description algorithm is applied to transform the extracted image information around feature points into low dimensional vector representation and store it. The vector representation obtained in this process is the image feature descriptor. Descriptors only pay attention to the part of information interested in the description algorithm and ignore the redundant information with low information entropy. At the same time, a low-dimensional vector or matrix can greatly save storage space.
The traditional image feature descriptors are divided into two types, one is based on local gradient and the other is based on image intensity. Image feature descriptor based on the local gradient is constructed by calculating and counting the histogram of gradient direction of the local area of an image. Local gradient image feature descriptor represented by SIFT (Scale Invariant Feature Transform) has been widely used in many image-based fields, such as image processing, computer vision, and so on.
SIFT is a landmark algorithm in the research of image local features. It was first proposed by David Lowe in 1999 and perfected in 2004 [23], which can detect and describe local features in images, find extreme points in scale space, and extract their position, scale, and rotation invariants. The core steps of SIFT feature extraction and description algorithm are divided into four steps, as showed in Figure 4. After two steps of Gaussian fuzzy processing, the image is input to scale space for extremum detection. The extremum points are mathematically processed to locate the key points. The image gradient method is used to determine the stable direction of the local structure. Finally, a group of vectors is used to describe image feature position, scale, and direction information. to eight quadrant circle grid partition, which enhances the robustness and independence of the descriptor. GLOH algorithm is based on log-polar coordinates when describing feature points. Three zones (6, 11, 15) and 8 angular directions are established in the radial direction in logarithmic polar coordinates. The detected region of interest is divided into 17 sub-regions according to the radial and angle. Then each region is divided into 16 bin gradients according to the SIFT descriptor method. The gradient angle is divided into 16 bins, and then the gradient histograms of each sub-region are spliced into a vector to get a 272-dimensional vector, namely 16 * 17. PCA is used to reduce the 272-dimensional vector to 40 dimensions.
SURF (speeded up robust features) is another improvement on SIFT proposed by Herbert et al. in 2008 [26]. Compared with SIFT, it is faster in execution. SURF uses the idea of simplified approximation in SIFT to simplify the Gauss second-order differential template in DOH so that only a few simple additions and subtraction operations are needed for the template to filter the image, and this operation is independent of the scale of the filter. Without down-sampling, the scale pyramid is constructed by keeping the image size unchanged but changing the size of the box filter, as showed in Figure 5. In the method of calculating the main direction of key points and the direction of pixels around the key points, SURF does not use histogram statistics but uses Haar wavelet transform. Relja et al. proposed RootSIFT [27] in 2012. Aiming at the similarity measurement method of SIFT features, a more accurate measurement method than Euclidean distance was proposed. After the description vector X of SIFT was extracted, the feature vector X was normalized by L1, and the square root of each element after normalization was calculated.
Another important descriptor is the feature descriptor based on image intensity. Since the LBP descriptors proposed by Ojala et al. in 1996, these descriptors have been improved and innovated and have been widely used. The basic idea is to express the image intensity difference in the local area around the feature points with binary string, then complete the image description and subsequent image matching in Hamming space [28].
LBP descriptors are often used to describe local texture features of images and are widely used in face recognition, facial expression recognition, and other fields. The basic LBP is defined in a 3 * 3 pixel neighborhood, the gray value of the center pixel is taken as the threshold value, as is shown in Figure 6. Then the gray value of the remaining 8 pixels in the neighborhood is compared with the threshold value, and the binary 1 and 0 are determined according to the comparison results. By connecting the eight binaries in order, we can get an 8-bit binary number, which is the LBP descriptor of a pixel. In 2002, Ojala et al. improved the LBP descriptors on scale and rotation invariance [29]. There are also many improvements on LBP feature descriptors, such as MB-LBP (Multi-Block LBP) [30], SEMB-LBP (Statistically Effective MB- LBP) [31]. LBP feature descriptors are characterized by low dimension, fast speed, rotation, and scale invariance. conditions. Therefore, in addition to the rotation and scale invariance, how to make the descriptors more robust to illumination changes should be considered when improving and innovating such descriptors. 3.3.
Machine learning and deep learning have developed rapidly in recent years, and are widely used in image processing. Especially after 2010, the combination of machine learning and traditional descriptors and learning methods to improve the extraction performance have become one of the hot directions of descriptors in recent years. Abandoning the traditional image feature descriptors algorithm flow, and finding a new way, the method of directly processing the image through deep learning to get the visual information description opens up a new direction of image processing. This method does not detect the feature points alone, but adaptively finds the feature description of the image according to the task by gradient descent according to the network structure and loss function.
Among the networks used in deep learning, the typical one is CNN (Convolutional Neural Networks). CNN is a kind of feedforward neural network that contains convolution computation and deep structure. It is one of the representative algorithms of deep learning. CNN has the ability of representation learning, which can classify the input information according to its hierarchical structure. Based on the basic model of CNN, LeNet5 marks the real appearance of CNN [36]; AlexNet is the first model that introduces convolutional neural networks into the field of computer vision and achieves breakthrough results [37]. Visual Geometry Group of Oxford proposed VGGNet and won the ILSVRC 2014 competition champion with a 25.3% error rate [38]. The 16-layer VGG version is shown in the figure 10, The concatenation of multiple convolutional blocks can effectively extract image features while making the network deeper. components are formed into a separable network by using spatial converter and soft-argmax function. Effective training methods and difficult sample mining strategies are designed specifically for this network to improve the discrimination of samples. In addition to the combination of depth network and feature point processing, Tateno et al. proposed to use a neural network to predict the depth map to carry out accurate and dense monocular reconstruction in 2017, naturally fusing the dense depth map predicted by CNN with the depth data measured by direct monocular SLAM [42]. In the fusion scheme, the depth prediction has priority in the image location with the large error of monocular SLAM (such as low texture area), while the priority of depth prediction is reduced in the image location with complex texture, which relies on more sufficient image feature descriptors for monocular reconstruction. The introduction of a depth network improves the accuracy and repeatability of monocular reconstruction. In addition, the combination of deep learning and traditional descriptors can also achieve adaptive feature extraction [43]. In 2018, Detone et al. proposed a full convolution neural network structure for interest point detection and description. The structure introduced the homographic adaptation method (a multi-scale, multi-source method that can improve the stability of interest point detection), and transformed sparse interest point detection and description into an efficient adaptive convolution neural network. Also in 2018, Ruihao Li et al. Proposed a new monocular unsupervised deep learning system UnDeepVO [44]. The system uses space loss and time loss between stereo image sequences for unsupervised training. During the test, the system can estimate the attitude and dense depth map of monocular images. This system is different from other model-based or learning-based monocular virtual reality methods. The scale restored in the training phase improves the recovery accuracy. The VO method based on unsupervised learning has the potential to improve performance with the increase of training data set. In addition, many types of research combine deep learning with traditional descriptors. They fully combine the advantages of deep learning automatic extraction with a fine-grained description of traditional descriptors and provide new ideas for the description of feature points.
The learning-based feature descriptor can be understood as the features collected from each layer of the network. Compared with the traditional descriptors, this is variable and more adaptive. For CNN, the feature descriptor is extracted in the form of the sliding window by using the convolution kernel learned, and the feature descriptor of the image is obtained by some activation functions such as Relu. The difference of characteristics between different layers depends on the scope of the receptive field and also on the task. It was mentioned in [45] that in the object detection task, the color information is the most felt by the bottom receptive field, and the number of perception units for the object will increase significantly with the increase of the number of layers. In the scene recognition task, the bottom receptive field has a stronger perception of objects and some areas of objects. When the number of layers rises, it will focus on the color of the scene. Each layer in the network extracts a feature descriptor, and the latter extracts and integrates from the previous descriptors to get a better embedding, which gradually approaches the feature space of ground truth. Therefore, from the bottom layer of the network to the high level of the network, the extracted descriptor will be more suitable for tasks.
There are several innovative directions for researchers' reference. When extracting feature points, It can be chosen whether to use an image scale pyramid or not according to whether scale invariance is needed, as well as the feature direction for rotation invariance. In addition, the distribution of feature points should be more uniform to avoid the direct influence of partial feature disappearance and shadow caused by illumination condition change on image matching. The descriptor is a sparse image representation. The dimension of the image feature descriptor is related to storage space and matching speed. The feature dimension or the number of stored feature descriptors should be reduced under the demands of the matching performance. The detected feature points and descriptors can be integrated according to the matching performance to avoid repeated descriptions.
SIFT descriptors and various optimized descriptors derived from SIFT have been widely used in various image-based algorithmic scenarios because of their scale, rotation invariance, and the advantages of effectively suppressing the influence of illumination and noise. Image mosaic technology is needed in computer vision, virtual reality, remote sensing image processing, video monitoring, and other fields. The most important role of image mosaic is image feature descriptor. The quality and efficiency of mosaic will be directly determined by the quality of feature extraction and description algorithm. SIFT algorithm is recognized as a more effective and stable method because of its advantages [46]. As one of the key technologies in digital image processing, image registration technology has become the core foundation of image mosaic [47] and object detection and tracking [48]. Image registration is the process of spatial matching of two or more images collected by different sensors under different environmental conditions. SIFT descriptors are often used in feature descriptions of image registration [49]. It is recognized as a more effective and stable method. The main conventional algorithm used in image retrieval and classification is BoW [50]. The core idea is to extract the key point descriptors and train a codebook by the clustering method. Then the number of times each descriptor vector in each image appears in the codebook is used to represent the image. SIFT descriptor is often used in the keypoint descriptors.
In the field of indoor visual positioning, the feature descriptors play an important role in many kinds of algorithm models. For example, the hierarchical structure position estimation algorithm [51] proposed by SMILE Lab of School of Electronics and Information Engineering, Xi'an Jiaotong University in 2013 used SIFT to represent the local features of the images in the dataset, and the SIFT obtained are used for subsequent clustering work. In 2015, the Multimedia Computing Group of Delft University of Technology proposed the Geo-Visual Ranking [52], which only uses visual images to predict geographic coordinates. The candidate image selection in the model relies on SIFT descriptor for image characterization and similarity measurement. In 2015, Deretey et al. Proposed a method for indoor positioning using a monocular camera [53]. In the process of generating a 3D map in the offline phase, the SIFT / SURF descriptors are used for 2D-3D feature projection.
Binary descriptors are widely used in face recognition, object recognition, SLAM (Simultaneous Localization And Mapping) scenes because of their advantages such as relatively simple calculation, low feature dimension, and fast matching speed. The system of positioning and navigation based on binary descriptors complements other data sources and achieves good results. Qian et al. Proposed a fast face recognition system based on PLBP (Pyramid of Local Binary Pattern) [54] and RI-LBP (Rotation Invariant Local Binary Pattern) [55]. Hussain et al. Developed an LQP (Local Quantified Pattern) technology for face recognition [56].
The advantages of fast speed, low dimension, and wide applicability of the binary descriptor also make it popular in the direction of visual positioning. Qin et al. developed a set of high-quality visualinertial navigation positioning systems, namely VINS-MONO, and carried out open-source sharing [57]. VINS uses BRIEF as the descriptor of the system's visual positioning and loop detection. After the features are extracted by the FAST feature extraction algorithm, the BRIEF feature description algorithm is used to describe and save the obtained descriptor, and the sparse point cloud is constructed according to the monocular camera. Through the mapping relationship between the matching features of the two images, the motion of the robot is calculated when images are acquired. At the same time, the feature descriptor point cloud map is constructed to judge whether the robot has been to its current position. This kind of loop detection technology is very important to correct the positioning error. The loop detection makes good use of the advantages of the low dimension, easy storage, space-saving, and fast matching speed of BRIEF, which ensures the real-time positioning to a certain extent. ORB is also popular in applications with high requirements of real-time and stability, such as the ORB-SLAM system with better implementation performance in the SLAM field [58]. Based on the ORB feature detection and description algorithm, this system builds a set of an indoor positioning systems for realtime positioning and mapping. The positioning and pose optimization speed is fast, the ORB sparse point cloud map is stable, and the comprehensive accuracy, speed, and resource utilization are better than a dense map. Different from VINS, which uses an inertial navigation algorithm to assist positioning, the ORB-SLAM system's positioning process relies only on visual information, which makes it more difficult. VINS uses the data of the inertial measurement unit to calculate the rotation of the robot, and then uses the obtained rotation information for the extraction of visual information. Therefore, VINS only needs to use the BRIEF descriptor to achieve good results. ORB-SLAM is to introduce rotation information from the feature extraction stage and rotate the local image according to the rotation information in the description stage. It can achieve better results without assistance and makes the visual positioning results more robust to rotation. ORB-SLAM constructs the feature descriptor point cloud map for loop detection. The positioning error is corrected by matching the currently extracted feature descriptor with the descriptor stored in the database, to further improve the positioning accuracy. The matching speed of ORB feature descriptors of the same dimension is the same as that of BRIEF, so the real-time positioning can be guaranteed. Leutenegger et.al. developed a visual-inertial navigation positioning system based on the BRISK descriptor in 2013, namely OKVIS [59]. This is a real-time positioning system using a stereo camera and IMU. After using the Harris corner detection algorithm for feature detection, BRISK is used for feature description. In positioning, the IMU is used to obtain the predicted value of the "next position". During feature matching, feature matching is performed between the current frame and the previous frame. Between the current frame and the image frame of the predicted position, the feature descriptors that can be matched are selected. The mapping relationship between the feature points in the 3D space and the image is established through the image matching results. In feature matching, instead of guiding the matching region before matching, OKVIS directly uses violence to find the descriptors that can complete the matching in the selected descriptors, which can also ensure the real-time positioning to a certain extent. However, OKVIS has no closedloop detection or correction scheme, so it is easy to accumulate positioning drift after long-time or longdistance movement, which reduces the accuracy of positioning results. Moreover, in terms of CPU utilization, OKVIS using Harris corner detection algorithm and BRISK is higher than VINS-MONO which uses FAST and BRIEF.
The computer vision field based on the CNN algorithm derives many subdivision application scenarios, such as camera relocation, target detection, instance segmentation, scene recognition, depth estimation, image alignment, etc. There may be more subdivisions based on different tasks, such as vehicle re-identification, pedestrian detection, and other tasks in the instance segmentation field.
Indoor positioning involves camera relocation, image retrieval, and other fields. The camera relocation task includes the prediction of six degrees of freedom of the camera: the position and height of the camera, and the rotation angle of three dimensions. RelocNet proposes a convolutional network representation learning method for camera pose retrieval [60], which is based on feature descriptors of nearest neighbor matching and continuous metric learning. The feature embedding network is optimized by using the overlap information of camera cones between image pairs. The difference between the final camera pose descriptors of the network represents the change of camera pose. In addition, they also constructed a pose regression, which uses geometric loss to train, and obtain a more precise relative pose relationship between images based on query image and nearest neighbor image. Experimental results show that the proposed method can be generalized in different datasets. In 7Scenes and RelocDB datasets, the matching success rate is about 70%, which is better than other methods. Based on the image retrieval method, the two-dimensional plane position information of the camera is obtained by comparing the image with the fingerprint collected from different places. The input image can be a picture in the same direction, a panoramic picture, or a picture based on some specific objects. The image data processed by [61] collect images from four directions (front, back, left, and right). After the fingerprint map of the network is established, the input image will use the features extracted by the convolution network, and then input into Graph Location Networks for positioning. Finally, the matching accuracy reaches 93.92% in ICUBE.
Image feature descriptor is one of the key technologies for image matching, and it has been widely used in many fields. In the field of localization, a trade-off should be made between the speed of the image feature descriptor and the robustness of different conditions. The image feature descriptor which is more suitable for the task should be selected. The innovative image feature descriptor should be designed according to the needs. In addition, relying solely on image feature descriptors for indoor positioning has many difficulties to overcome, and the performance might be not so satisfactory. More and more researchers tend to use multi-sensor fusion for localization, which also provides a new direction and possibility for the expansion and application of image feature descriptors.
In this review, we conduct a comprehensive overview of the various image feature descriptors that may be used in visual positioning. For each kind of descriptor, we summarize the principle and improvement of the algorithm in detail. We provide a thorough review, comparisons, and summarizations of descriptors and their applications in recent years. Finally, we summarize the applications of features descriptors in image matching or indoor visual positioning. We hope this paper can provide innovative ideas and references for researchers.
This work was financially supported by the National Natural Science Foundation of China (No.61871054).
[20] Lowe, D.G. Object recognition from local scale-invariant features. Proceedings of the International
Conference on Computer Vision 1999, 1150-1157. [21] Rosten E . Machine learning for high-speed corner detection. European Conference on Computer
Vision. Springer-Verlag 2006. [22] Khwildi R , Azza Ouled Zaid. HDR image retrieval by using color-based descriptor and tone mapping operator. The Visual Computer 2019, (8). [23] Lowe D G . Distinctive image features from scale-invariant keypoints. International Journal of
Computer Vision 2004, 60(2):91-110. [24] Yan Ke, Rahul Sukthankar. PCA-SIFT: A more distinctive representation for local image descriptors. Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’04) 2004, 1063-6919/04. [25] Krystian Mikolajczyk and Cordelia Schmid "A performance evaluation of local descriptors". IEEE
Transactions on Pattern Analysis and Machine Intelligence 2005, 10, 27, pp 1615--1630. [26] Herbert Bay , Andreas Ess, "SURF: Speeded up robust features". Computer Vision and Image
Understanding (CVIU) 2008, Vol. 110, No. 3, pp. 346–359. [27] Relja Arandjelović; Andrew Zisserman, Three things everyone should know to improve object retrieval. IEEE Conference on Computer Vision and Pattern Recognition 2012, 16-21 June. [28] Yongqiang G , Yu Q , Weilin H . Local binary descriptors and its application to image matching.
Journal of Network New Media 2014. [29] Ojala T , Pietikainen M , Maenpaa T . Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE 2002:971-987. [30] Liao S C , Zhu X X , Lei Z , et al. Learning multi-scale block local binary patterns for face recognition. In Lecture Notes in Computer Science 4642; Center for Biometrics and Security Research & National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, 95 Zhongguancun Donglu, Beijing 100080, China, 2007. [31] Ebadi S , Ansari N N , Naghdi S , et al. The effect of continuous ultrasound on chronic non-specific low back pain: a single blind placebo-controlled randomized trial. Bmc Musculoskeletal Disorders 2012, 13. [32] Calonder M , Lepetit V , Strecha C , et al. BRIEF: Binary robust independent elementary features.
Computer Vision - ECCV 2010, 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV. Springer, Berlin, Heidelberg 2010. [33] Rublee E , Rabaud V , Konolige K , et al. ORB: An efficient alternative to SIFT or SURF.
International Conference on Computer Vision. IEEE 2012. [34] Leutenegger S , Chli M , Siegwart R Y . BRISK: Binary robust invariant scalable keypoints.
International Conference on Computer Vision. IEEE 2011. [35] Alahi A , Ortiz R , Vandergheynst P . FREAK: Fast retina keypoint. IEEE Conference on Computer
Vision & Pattern Recognition. IEEE 2012. [36] El-Sawy A, Hazem E L B, Loey M. CNN for handwritten arabic digits recognition based on LeNet5. International conference on advanced intelligent systems and informatics. Springer, Cham 2016: 566-575. [37] Krizhevsky A , Sutskever I , Hinton G . ImageNet classification with deep convolutional neural networks. NIPS. Curran Associates Inc 2012. [38] Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition.
arXiv preprint 2014:1409.1556. [39] Veit A, Wilber M J, Belongie S. Residual networks behave like ensembles of relatively shallow networks. Advances in neural information processing systems 2016: 550-558. [40] Tan M, Le Q. Efficientnet: Rethinking model scaling for convolutional neural networks[C]//International Conference on Machine Learning. PMLR, 2019: 6105-6114. [41] Yi K M , Trulls E , Lepetit V , et al. LIFT: Learned invariant seature transform. European
Conference on Computer Vision. Springer, Cham 2016. [42] Tateno K , Tombari F , Laina I , et al. CNN-SLAM: Real-time dense monocular SLAM with learned depth prediction. Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society 2017. [43] D. Detone, T. Malisiewicz, and A. Rabinovich. Superpoint: Self-supervised interest point detection and description. CVPR Workshop on Deep Learning for Visual SLAM 2018. [44] R. Li, S. Wang, Z. Long, and D. Gu, “UnDeepVO: Monocular visual odometry through unsupervised deep learning,”. arXiv preprint arXiv 2017:1709.06841. [45] Bau D, Zhu J Y, Strobelt H, et al. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences 2020. [46] LUO J,GWUN O. A comparison of SIFT, PCA-SIFT and SURF. Inernational J. Image Processing (LP) 2009, 3(4) :143-152. [47] Zhang F,Liu F.Parallax-tolerant image stitching. 2014IEEE Conference on Computer Vision and
Pattern Recognition(CVPR).IEEE 2014:3262-3269. [48] Barbara Zitova, Flusser J.Image registration methods: a survey. Image and Vision Computing 2003,21(11):977-1000. [49] Adel E,Elmogy M,Elbakry H. Image stitching based on feature extraction techniques:a survey.
International Journal of Computer Applications 2014, 99(6):1-8. [50] HervéJégou, Douze M , Schmid C . Packing bag-of-features. EEE International Conference on
Computer Vision. IEEE 2010. [51] Li J , Qian X , Tang Y Y , et al. GPS estimation for places of interest from social users' uploaded photos. IEEE Transactions on Multimedia 2013, 15(8):2058-2071. [52] Li, Xinchao, Larson, et al. Global-scale location prediction for social images using geo-visual ranking. IEEE Transactions on Multimedia 2015. [53] Deretey E , Ahmed M T , Marshall J A , et al. Visual indoor positioning with a single camera using PnP. 2015 International Conference on Indoor Positioning and Indoor Navigation (IPIN). IEEE 2016. [54] Qian X , Hua X S , Chen P , et al. PLBP: An effective local binary patterns texture descriptor with pyramid representation. Pattern Recognition 2011,44(10-11):2502-2515. [55] Guo Z , Zhang L , Zhang D . Rotation invariant texture classification using LBP variance (LBPV) with global matching. Pattern Recognition 2010, 43( 3):706-719. [56] Hussain, S.; Triggs, B. Visual recognition using local quantized patterns. In Computer Vision
ECCV 2012; pp. 716–729. [57] Qin T, Li P, Shen S. Vins-mono: A robust and versatile monocular visual-inertial state estimator.
IEEE Transactions on Robotics 2018, 34(4): 1004-1020. [58] Mur-Artal R, Montiel J M M, Tardos J D. ORB-SLAM: a versatile and accurate monocular SLAM system. IEEE transactions on robotics 2015, 31(5): 1147-1163.. [59] Leutenegger S , Furgale P , Rabaud V , et al. Keyframe-based visual-inertial SLAM using nonlinear optimization. Proceedings of Robotics: Science and Systems 2013. [60] Balntas V, Li S, Prisacariu V. Relocnet: Continuous metric learning relocalisation using neural nets. Proceedings of the European Conference on Computer Vision (ECCV) 2018: 751-767. [61] Chiou M J, Liu Z, Yin Y, et al. Zero-shot multi-view indoor localization via graph location networks. Proceedings of the 28th ACM International Conference on Multimedia. 2020: 34313440.