DOCSLIB.ORG

Omnidirectional Visual-Inertial Odometry Using Multi-State Constraint Kalman Filter

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Omnidirectional Visual-Inertial Odometry Using Multi-State Constraint Kalman Filter Omnidirectional Visual-Inertial Odometry Using Multi-State Constraint Kalman Filter Milad Ramezani 1 and Kourosh Khoshelham 1 and Laurent Kneip 2 Abstract— We present an Omnidirectional Visual-Inertial VIO can be substantially improved by increasing the Field of Odometry (OVIO) approach based on Multi-State Constraint View (FoV) of the camera. A larger FoV allows capturing Kalman Filtering (MSCKF) to estimate the ego-motion of a a larger extent of the environment, thereby increasing the moving platform. Instead of considering visual measurements on image plane, we use individual planes for each point that likelihood of usable regions in images for feature extraction are tangent to the unit sphere and normal to the corresponding and matching. Furthermore, a larger FoV provides a better measurement ray. This way, we combine spherical images cap- geometry for the pose estimation due to the fact that it tured by omnidirectional camera with inertial measurements is more likely to track features from different angles sur- within the filtering method MSCKF. The key hypothesis of rounding the moving camera. More importantly, a wider OVIO is that a wider field of view allows incorporating more visual features from the surrounding environment, thereby im- FoV allows features to be tracked over a longer sequence proving the accuracy and robustness of the motion estimation. of images. In theory, this increases the number of visual Moreover, by using an omnidirectional camera, it is less likely measurements, and consequently improves the precision and to end up in a situation where there is not enough texture. robustness of pose estimation. It is worth noting that larger We provide an evaluation of OVIO using synthetic and real FoV can solve the rotation- translation ambiguity which is a video sequences captured by a fish-eye camera, and compare the performance with MSCKF using a perspective camera. The serious problem with smaller FoVs. results show the superior performance of the proposed OVIO. Despite these clear advantages, until now only little at- tention has been paid to large FoV and omnidirectional I. INTRODUCTION cameras for visual inertial odometry. To fill this gap, this Pose estimation by means of visual sensors is becoming paper presents an Omnidirectional Visual-Inertial Odometry important in many applications including navigation and (OVIO) approach using Multi-State Constraint Kalman Fil- autonomous vehicles. The demand for accurate position tering (MSCKF) [5], which integrates visual information ex- information is more pronounced in areas where the Global tracted from an omnidirectional camera with the inertial mea- Positioning System (GPS) has no coverage, such as urban surements from an INS. Since the standard perspective model canyons. is not suitable for larger FoVs, we develop a measurement A common approach to visual navigation is Visual Odom- model in MSCKF that incorporates measurement residuals etry (VO), which uses only visual information to estimate on the unit sphere rather than on the image plane. Moreover, motion. This method is highly sensitive to the quality of we apply an omnidirectional model to calculate the global image observations. By combining visual features from a 3D position of features and propagate their uncertainties. video camera with inertial measurements from an Inertial We evaluate OVIO and compare its performance with the Navigation System (INS), the resilience in low-texture sit- conventional VIO within MSCKF through experiments in uations and accuracy can be substantially improved. This synthetic and real environments. technique, which is known as Visual-Inertial Odometry The paper is organized as follows: After a brief discussion (VIO), has been an active research area for more than a of the state-of-the-art VIO algorithms in section II, we intro- decade [1]–[4]. VIO employs visual measurements in order duce in section III our Omnidirectional Visual Inertial Odom- to compensate for INS errors within an estimation model. etry (OVIO) approach using MSCKF for pose estimation. In Today, most VIO systems use visual information from a section IV, we evaluate our approach and compare it with perspective camera with a limited Field of View (FoV). MSCKF using a perspective camera through experiments Therefore, achieving a robust positioning performance by in synthetic and real environments. Finally, conclusions are the state-of-the-art VIO approaches is still a challenge in drawn and directions for future research are discussed. environments that lack salient features due to illumination changes or insufficient texture in the entire image as well as II. RALATED WORK the presence of moving objects covering a significant area Algorithms for the integration of an exteroceptive sensor of the image of a perspective camera. The robustness of such as a camera with a proprioceptive inertial sensor are 1M. Ramezani and K. Khoshelham are with the Department known as Visual-Inertial Odometry (VIO) and have been of Infrastructure Engineering, The University of Melbourne, investigated through a vast body of literature in robotics [1]- Australia [email protected], [5]. From the estimation point of view, VIO approaches can [email protected] 2L. Kneip is with the ShanghaiTech School of Information Science and be classified into optimization-based and filtering methods. Technology, Shanghai, China [email protected] In optimization methods, unknown parameters such as the sensor pose are estimated by iterative minimization of a least- In the last few years these cameras have gained popularity squares error function. In [6] the sensor pose, scene structure, in robotics [23], [25], and a substantial body of work has and other unknowns are estimated using all image and been devoted to modelling their imaging geometry [26], inertial measurements simultaneously. However, because this [27] and calibration [28]–[30]. Despite the advantages of method requires that all measurements from all images are omnidirectional cameras for ego-motion estimation, they available before the estimation, it is not suitable for real-time have been mostly used in VO [23]–[25], but rarely in VIO. applications. In order to solve the real-time operability issue, In [31] an effort was made in order to fuse omnidirectional Leutenegger et al. [7] bounded the size of the optimization vision with inertial sensors using an adaptive estimator. This window by marginalizing out old states. This method is method estimates the position and linear velocity of a mobile introduced as Sliding Window Filter (SWF) in [8]. The robot, but not other pose parameters such as orientation. shortcoming of this approach is that not all measurements of Furthermore, the performance of the algorithm has not been a seen feature are used, because multiple poses are discarded tested in a large scale environment. resulting in a loss of constraints. Overall the review of the literature shows a preference for In contrast to non-linear optimization, filtering methods perspective cameras in VO and VIO approaches. Motivated estimate sensor motion and other parameters recursively in a by the potential of omnidirectional cameras for pose esti- filter as measurements from INS and camera become avail- mation, this work presents an omnidirectional visual-inertial able. Therefore, this approach is appropriate for real-time odometry (OVIO) approach. applications even in large-scale environments with infinitely III. OMNIDIRECTIONAL VISUAL-INERTIAL long image sequences [6]. A popular estimator in filtering ODOMETRY methods is the Extended Kalman Filter (EKF) [1], [2], [9]– [11]. In EKF-based methods, a state vector that comprises We begin by briefly describing the system model, and then the unknowns is first propagated in a linear system model introduce a new measurement model in the MSCKF based on and then is updated by feature measurements extracted from the tangential plane around the measurement ray on the unit the images. If the scene structure and the current INS pose sphere. After defining the new measurement model, we can are simultaneously estimated, the approach is called EKF- link the omnidirectional observations with the system model based Simultaneous Localization and Mapping (SLAM) to update the state vector and its covariance. A detailed [12]–[15]. An advantage of EKF-SLAM algorithms is that description of the MSCKF can be found in [1], [5]. the correlations between camera motion and 3D features are A. System model considered since camera poses and 3D positions of observed M × features are estimated together. However, correctly handling Let X be the (16+7 ) 1 EKF state vector defined as: these correlations imposes computational complexity that BT T E T T E T compromises this technique in an environment with many X = [qE bg vB ba pB ··· qC1T EpT ··· qCiT EpT ··· qCM T EpT ]T features [1]. E C1 E Ci E CM There are several filtering algorithms in which, contrary to (1) B SLAM, the global position of features is not jointly estimated where qE is the 4 × 1 unit quaternion representing the in the state vector [5], [16]–[21]. These methods utilize the orientation of the INS resolved in the body frame fBg, E E visual measurements to impose epipolar constraints between vectors vB, and pB are the linear velocity and the either pairs of images [16]–[18] or multiple camera poses [5], position of the INS w.r.t. the Earth frame fEg, respectively, [22]. Amongst these methods, Multi-State Constraint Kalman bg and ba are the 3 × 1 vectors describing the gyroscope and Filtering (MSCKF) [5] benefits from the entire geometric accelerometer biases, respectively. The remaining parameters Ci constraints that a feature produces when seen from multiple are M camera poses that include the camera quaternions, qE , camera poses. By employing all these geometric constraints, relative to the camera frame fCig, i = 1 ··· M and positions, E fEg more accurate estimates of the camera poses can be obtained. pCi ,in the Earth frame . Following equation (1), the What all vision-aided navigation systems have in common (15 + 6M) × 1 error-state vector is: is that they are highly dependent on visual information observed from the environment.
Load more

Recommended publications
  • Semantically Informed Visual Odometry and Mapping
    SIVO: Semantically Informed Visual Odometry and Mapping by Pranav Ganti A thesis presented to the University of Waterloo in fulfillment of the thesis requirement for the degree of Master of Applied Science in Mechanical and Mechatronics Engineering Waterloo, Ontario, Canada, 2018 c Pranav Ganti 2018 I hereby declare that I am the sole author of this thesis. This is a true copy of the thesis, including any required final revisions, as accepted by my examiners. I understand that my thesis may be made electronically available to the public. ii Abstract Accurate localization is a requirement for any autonomous mobile robot. In recent years, cameras have proven to be a reliable, cheap, and effective sensor to achieve this goal. Visual simultaneous localization and mapping (SLAM) algorithms determine camera motion by tracking the motion of reference points from the scene. However, these references must be static, as well as viewpoint, scale, and rotation invariant in order to ensure accurate localization. This is especially paramount for long-term robot operation, where we require our references to be stable over long durations and also require careful point selection to maintain the runtime and storage complexity of the algorithm while the robot navigates through its environment. In this thesis, we present SIVO (Semantically Informed Visual Odometry and Mapping), a novel feature selection method for visual SLAM which incorporates machine learning and neural network uncertainty into an information-theoretic approach to feature selection. The emergence of deep learning techniques has resulted in remarkable advances in scene understanding, and our method supplements traditional visual SLAM with this contextual knowledge.
    [Show full text]
  • Visual Odometry by Multi-Frame Feature Integration
    Visual Odometry by Multi-frame Feature Integration Hernan´ Badino Akihiro Yamamoto∗ Takeo Kanade [email protected] [email protected] [email protected] Robotics Institute Carnegie Mellon University Pittsburgh, PA Abstract Our proposed method follows the procedure of comput- ing optical flow and stereo disparity to minimize the re- This paper presents a novel stereo-based visual odom- projection error of tracked feature points. However, instead etry approach that provides state-of-the-art results in real of performing this task using only consecutive frames (e.g., time, both indoors and outdoors. Our proposed method [3, 21]), we propose a novel and computationally simple follows the procedure of computing optical flow and stereo technique that uses the whole history of the tracked features disparity to minimize the re-projection error of tracked fea- to compute the motion of the camera. In our technique, ture points. However, instead of following the traditional the features measured and tracked over all past frames are approach of performing this task using only consecutive integrated into a single improved estimate of the real 3D frames, we propose a novel and computationally inexpen- feature projection. In this context, the term multi-frame fea- sive technique that uses the whole history of the tracked ture integration refers to this estimation and noise reduction feature points to compute the motion of the camera. In technique. our technique, which we call multi-frame feature integra- This paper presents three main contributions: tion, the features measured and tracked over all past frames are integrated into a single, improved estimate.
    [Show full text]
  • Spherical-Model-Based SLAM on Full-View Images for Indoor Environments
    applied sciences Article Spherical-Model-Based SLAM on Full-View Images for Indoor Environments Jianfeng Li 1 , Xiaowei Wang 2 and Shigang Li 1,2,* 1 School of Electronic and Information Engineering, also with the Key Laboratory of Non-linear Circuit and Intelligent Information Processing, Southwest University, Chongqing 400715, China; [email protected] 2 Graduate School of Information Sciences, Hiroshima City University, Hiroshima 7313194, Japan; [email protected] * Correspondence: [email protected] Received: 14 October 2018; Accepted: 14 November 2018; Published: 16 November 2018 Featured Application: For an indoor mobile robot, finding its location and building environment maps. Abstract: As we know, SLAM (Simultaneous Localization and Mapping) relies on surroundings. A full-view image provides more benefits to SLAM than a limited-view image. In this paper, we present a spherical-model-based SLAM on full-view images for indoor environments. Unlike traditional limited-view images, the full-view image has its own specific imaging principle (which is nonlinear), and is accompanied by distortions. Thus, specific techniques are needed for processing a full-view image. In the proposed method, we first use a spherical model to express the full-view image. Then, the algorithms are implemented based on the spherical model, including feature points extraction, feature points matching, 2D-3D connection, and projection and back-projection of scene points. Thanks to the full field of view, the experiments show that the proposed method effectively handles sparse-feature or partially non-feature environments, and also achieves high accuracy in localization and mapping. An experiment is conducted to prove that the accuracy is affected by the view field.
    [Show full text]
  • Evolution of Visual Odometry Techniques
    Evolution of Visual Odometry Techniques Shashi Poddar, Rahul Kottath, Vinod Karar Abstract— With rapid advancements in the area of mobile robotics visual odometry (VO) pipeline [4]. Simultaneous localization and industrial automation, a growing need has arisen towards accurate and mapping (SLAM), a superset of VO, localizes and builds navigation and localization of moving objects. Camera based motion a map of its environment along with the trajectory of a moving estimation is one such technique which is gaining huge popularity owing to its simplicity and use of limited resources in generating motion path. object [5]. However, our discussion in this paper is limited to In this paper, an attempt is made to introduce this topic for beginners visual odometry, which incrementally estimates the camera covering different aspects of vision based motion estimation task. The pose and refines it using optimization technique. A visual evolution of VO schemes over last few decades is discussed under two odometry system consists of a specific camera arrangement, broad categories, that is, geometric and non-geometric approaches. The geometric approaches are further detailed under three different classes, the software architecture and the hardware platform to yield that is, feature-based, appearance-based, and a hybrid of feature and camera pose at every time instant. The camera pose estimation appearance based schemes. The non-geometric approach is one of the can be either appearance or feature based. The appearance- recent paradigm shift from conventional pose estimation technique and based techniques operate on intensity values directly and is thus discussed in a separate section. Towards the end, a list of different datasets for visual odometry and allied research areas are provided for a matches template of sub-images over two frame or the optical ready reference.
    [Show full text]
  • Semi-Dense Visual Odometry for AR on a Smartphone
    Semi-Dense Visual Odometry for AR on a Smartphone Thomas Schops¨ ∗ Jakob Engely Daniel Cremersz Technische Universitat¨ Munchen¨ Figure 1: From left to right: AR demo application with simulated car. Corresponding estimated semi-dense depth map. Estimated dense collision mesh, fixed and shown from a different perspective. Photo of running system. The attached video shows the system in action. ABSTRACT of-the-art monocular SLAM methods generally operate on features. We present a direct monocular visual odometry system which runs While this allows to estimate the camera movement in real-time on in real-time on a smartphone. Being a direct method, it tracks and mobile platforms [12, 15], the resulting feature based maps hardly maps on the images themselves instead of extracted features such as provide sufficient information about the 3D geometry of the scene keypoints. New images are tracked using direct image alignment, for physical interaction. while geometry is represented in the form of a semi-dense depth At the same time, recent advances in computer vision have direct map. Depth is estimated by filtering over many small-baseline, shown the high potential of methods for monocular pixel-wise stereo comparisons. This leads to significantly less out- SLAM [16, 5, 7, 17]: instead of operating on features, these meth- liers and allows to map and use all image regions with sufficient ods perform both tracking and mapping directly on the image inten- gradient, including edges. We show how a simple world model sity values. Fundamentally different from feature based methods, for AR applications can be derived from semi-dense depth maps, direct methods not only allow for fast, sub-pixel accurate camera and demonstrate the practical applicability in the context of an AR tracking, but also provide substantially more information about the application in which simulated objects can collide with real geom- 3D structure of the environment, are less susceptible to outliers, and etry.
    [Show full text]
  • Optical Flow and Deep Learning Based Approach to Visual Odometry
    Rochester Institute of Technology RIT Scholar Works Theses 11-2016 Optical Flow and Deep Learning Based Approach to Visual Odometry Peter M. Muller [email protected] Follow this and additional works at: https://scholarworks.rit.edu/theses Recommended Citation Muller, Peter M., "Optical Flow and Deep Learning Based Approach to Visual Odometry" (2016). Thesis. Rochester Institute of Technology. Accessed from This Thesis is brought to you for free and open access by RIT Scholar Works. It has been accepted for inclusion in Theses by an authorized administrator of RIT Scholar Works. For more information, please contact [email protected]. Optical Flow and Deep Learning Based Approach to Visual Odometry Peter M. Muller Optical Flow and Deep Learning Based Approach to Visual Odometry Peter M. Muller November 2016 A Thesis Submitted in Partial Fulfillment of the Requirements for the Degree of Master of Science in Computer Engineering Department of Computer Engineering Optical Flow and Deep Learning Based Approach to Visual Odometry Peter M. Muller Committee Approval: Dr. Andreas Savakis Advisor Date Professor Dr. Raymond Ptucha Date Assistant Professor Dr. Roy Melton Date Principal Lecturer i Acknowledgments I would like to thank my advisor, Dr. Andreas Savakis, for his exceptional patience and direction in helping me complete this thesis. I would like to thank my committee, Dr. Raymond Ptucha and Dr. Roy Melton, for being outstanding mentors during my time at RIT. I would also like to thank my fellow lab member, Dan Chianucci, for his help and input during the course of this thesis. Finally, I would like to thank my friends who have supported me from distances near and far: anywhere from just beyond the lab doors to just beyond the other side of the world.
    [Show full text]
  • Unsupervised Deep Learning-Based RGB-D Visual Odometry
    Article Unsupervised Deep Learning-Based RGB-D Visual Odometry Qiang Liu1,2 , Haidong Zhang 2, Yiming Xu 2,* and Li Wang 2 1 Department of Psychiatry, University of Oxford, Warneford Hospital, Oxford OX3 7JX, UK; [email protected] 2 School of Electrical Engineering, Nantong University, Nantong 226000, China; [email protected] (H.Z.); [email protected] (L.W.) * Correspondence: [email protected] Received: 6 July 2020; Accepted: 3 August 2020; Published: 6 August 2020 Abstract: Recently, deep learning frameworks have been deployed in visual odometry systems and achieved comparable results to traditional feature matching based systems. However, most deep learning-based frameworks inevitably need labeled data as ground truth for training. On the other hand, monocular odometry systems are incapable of restoring absolute scale. External or prior information has to be introduced for scale recovery. To solve these problems, we present a novel deep learning-based RGB-D visual odometry system. Our two main contributions are: (i) during network training and pose estimation, the depth images are fed into the network to form a dual-stream structure with the RGB images, and a dual-stream deep neural network is proposed. (ii) the system adopts an unsupervised end-to-end training method, thus the labor-intensive data labeling task is not required. We have tested our system on the KITTI dataset, and results show that the proposed RGB-D Visual Odometry (VO) system has obvious advantages over other state-of-the-art systems in terms of both translation and rotation errors. Keywords: RGB-D sensor; visual odometry; unsupervised deep learning; Recurrent Convolutional Neural Networks 1.
    [Show full text]
  • Improved Ground-Based Monocular Visual Odometry Estimation Using Inertially-Aided Convolutional Neural Networks
    Air Force Institute of Technology AFIT Scholar Theses and Dissertations Student Graduate Works 3-2020 Improved Ground-Based Monocular Visual Odometry Estimation using Inertially-Aided Convolutional Neural Networks Josiah D. Watson Follow this and additional works at: https://scholar.afit.edu/etd Part of the Electrical and Electronics Commons, and the Navigation, Guidance, Control and Dynamics Commons Recommended Citation Watson, Josiah D., "Improved Ground-Based Monocular Visual Odometry Estimation using Inertially-Aided Convolutional Neural Networks" (2020). Theses and Dissertations. 3622. https://scholar.afit.edu/etd/3622 This Thesis is brought to you for free and open access by the Student Graduate Works at AFIT Scholar. It has been accepted for inclusion in Theses and Dissertations by an authorized administrator of AFIT Scholar. For more information, please contact [email protected]. IMPROVED GROUND-BASED MONOCULAR VISUAL ODOMETRY ESTIMATION USING INERTIALLY-AIDED CONVOLUTIONAL NEURAL NETWORKS THESIS Josiah D Watson AFIT-ENG-MS-20-M-072 DEPARTMENT OF THE AIR FORCE AIR UNIVERSITY AIR FORCE INSTITUTE OF TECHNOLOGY Wright-Patterson Air Force Base, Ohio DISTRIBUTION STATEMENT A APPROVED FOR PUBLIC RELEASE; DISTRIBUTION UNLIMITED. The views expressed in this document are those of the author and do not reflect the official policy or position of the United States Air Force, the United States Department of Defense or the United States Government. This material is declared a work of the U.S. Government and is not subject to copyright protection in the United States. AFIT-ENG-MS-20-M-072 Improved Ground-Based Monocular Visual Odometry Estimation using Inertially-Aided Convolutional Neural Networks THESIS Presented to the Faculty Department of Electrical and Computer Engineering Graduate School of Engineering and Management Air Force Institute of Technology Air University Air Education and Training Command in Partial Fulfillment of the Requirements for the Degree of Master of Science in Electrical Engineering Josiah D Watson, B.S.Cp.E.
    [Show full text]
  • Real-Time Stereo Visual Odometry for Autonomous Ground Vehicles
    Real-Time Stereo Visual Odometry for Autonomous Ground Vehicles Andrew Howard Abstract— This paper describes a visual odometry algorithm in the inlier set. The inlier detection step (3) is the key for estimating frame-to-frame camera motion from successive distinguishing feature of the algorithm. The feature matching stereo image pairs. The algorithm differs from most visual stage inevitably produces some incorrect correspondences, odometry algorithms in two key respects: (1) it makes no prior assumptions about camera motion, and (2) it operates on dense which, if left intact, will unfavorably bias the frame-to-frame disparity images computed by a separate stereo algorithm. motion estimate. A common solution to this problem is to This algorithm has been tested on many platforms, including use a robust estimator that can tolerate some number of false wheeled and legged vehicles, and has proven to be fast, accurate matches (e.g., RANSAC [4]). In our algorithm, however, and robust. For example, after 4000 frames and 400m of we adopt an approach described by Hirschmuller¨ et al. [7], travel, position errors are typically less than 1m (0.25% of distance traveled). Processing time is approximately 20ms on and exploit stereo range data at the inlier detection stage. a 512x384 image. This paper includes a detailed description The core intuition is that the 3D locations of features must of the algorithm and experimental evaluation on a variety of obey a rigidity constraint, and that this constraint can be platforms and terrain types. used to identify sets of features that are mutually consistent (i.e., a clique) prior to computing the frame-to-frame motion I.
    [Show full text]
  • Monocular Visual Odometry
    IIT-Kanpur Undergraduate Project 2 Monocular Visual Odometry Author: Supervisor: Avi Singh Dr. KS Venkatesh April 24, 2015 Contents 1 Introduction 2 1.1 Why Monocular? . 2 1.2 Application . 2 1.3 Related Work . 3 2 The Algorithm 3 2.1 Problem Formulation . 3 2.2 Algorithm Outline . 4 2.3 Undistortion . 4 2.4 Feature Detection . 4 2.5 Feature Tracking . 5 2.6 Essential Matrix Estimation . 5 2.7 Computing R, t from the Essential Matrix . 6 2.8 Constructing Trajectory . 6 2.9 Heuristics . 7 3 Results and Conclusions 7 4 Future Work 7 5 References 9 1 Abstract This report explains a monocular visual odometry algorithm, which has been implemented in OpenCV/C++. We report results on the KITTI dataset (using only one image of the stereo dataset). Since it is a monocular implementation, we cannot do absolute scale estima- tion, and thus that quantity is used from the ground truths that we have. 1 Introduction Visual Odometry is the estimation of 6-DOF trajectory followed by a mov- ing agent, based on input from a camera rigidly attached to the body of the agent. The camera might be monocular, or a couple of cameras might be used to form a stereo rig. In the monocular approach, it is not possible to obtain the absolute scale of the trajectory, while it is certainly possible to do so for the stereo approach, assuming we know the baseline (distance between the two cameras of the stereo rig). Visual Odometry has attracted a lot of research in the recent years, with new state-of-the-art approaches coming almost every year[14, 11].
    [Show full text]
  • Deep Direct Visual Odometry Chaoqiang Zhao, Yang Tang, Senior Member, IEEE, Qiyu Sun, and Athanasios V
    1 Deep Direct Visual Odometry Chaoqiang Zhao, Yang Tang, Senior Member, IEEE, Qiyu Sun, and Athanasios V. Vasilakos Abstract—Traditional monocular direct visual odometry (DVO) is one of the most famous methods to estimate the ego- motion of robots and map environments from images simul- taneously. However, DVO heavily relies on high-quality images and accurate initial pose estimation during tracking. With the outstanding performance of deep learning, previous works have shown that deep neural networks can effectively learn 6-DoF (Degree of Freedom) poses between frames from monocular image sequences in the unsupervised manner. However, these (a) DDSO on Seq. 10 unsupervised deep learning-based frameworks cannot accurately generate the full trajectory of a long monocular video because of the scale-inconsistency between each pose. To address this problem, we use several geometric constraints to improve the scale-consistency of the pose network, including improving the previous loss function and proposing a novel scale-to-trajectory constraint for unsupervised training. We call the pose network trained by the proposed novel constraint as TrajNet. In addition, a new DVO architecture, called deep direct sparse odometry (DDSO), is proposed to overcome the drawbacks of the previous direct sparse odometry (DSO) framework by embedding TrajNet. (b) DSO [7] on Seq. 10 Extensive experiments on the KITTI dataset show that the pro- posed constraints can effectively improve the scale-consistency of Fig. 1. Trajectories of DDSO and DSO [7] running on KITTI odometry TrajNet when compared with previous unsupervised monocular sequence 10. The proposed DDSO is more robust than DSO [7] in initial- methods, and integration with TrajNet makes the initialization ization.
    [Show full text]
  • VSO: Visual Semantic Odometry
    VSO: Visual Semantic Odometry Konstantinos-Nektarios Lianos 1,⋆, Johannes L. Sch¨onberger 2, Marc Pollefeys 2,3, Torsten Sattler 2 1 Geomagical Labs, Inc., USA 3 Microsoft, Switzerland 2 Department of Computer Science, ETH Z¨urich, Switzerland [email protected] {jsch,marc.pollefeys,sattlert}@inf.ethz.ch Abstract. Robust data association is a core problem of visual odometry, where image-to-image correspondences provide constraints for camera pose and map estimation. Current state-of-the-art direct and indirect methods use short-term tracking to obtain continuous frame-to-frame constraints, while long-term constraints are established using loop clo- sures. In this paper, we propose a novel visual semantic odometry (VSO) framework to enable medium-term continuous tracking of points using semantics. Our proposed framework can be easily integrated into exist- ing direct and indirect visual odometry pipelines. Experiments on chal- lenging real-world datasets demonstrate a significant improvement over state-of-the-art baselines in the context of autonomous driving simply by integrating our semantic constraints. Keywords: visual odometry, SLAM, semantic segmentation 1 Introduction Visual Odometry (VO) algorithms track the movement of one or multiple cam- eras using visual measurements. Their ability to determine the current position based on a camera feed forms a key component of any type of embodied artificial intelligence, e.g., self-driving cars or other autonomous robots, and of any type of intelligent augmentation system, e.g., Augmented or Mixed Reality devices. At its core, VO is a data association problem, as it establishes pixel-level associations between images. These correspondences are simultaneously used to build a 3D map of the scene and to track the pose of the current camera frame relative to the map.
    [Show full text]