SREF: Semantics-Refined Feature Extraction for Long-Term Visual Localization.

Visual localization has been well studied in recent decades and applied in many fields as a fundamental capability in robotics. However, the success of the state of the arts usually builds on the assumption that the environment is static. In dynamic scenarios where moving objects are present, the performance of the existing visual localization systems degrades a lot due to the disturbance of the dynamic factors. To address this problem, we propose a novel sparse motion removal (SMR) model that detects the dynamic and static regions for an input frame based on a Bayesian framework. The similarity between the consecutive frames and the difference between the current frame and the reference frame are both considered to reduce the detection uncertainty. After the detection process is finished, the dynamic regions are eliminated while the static ones are fed into a feature-based visual simultaneous localization and mapping (SLAM) system for further visual localization. To verify the proposed method, both qualitative and quantitative experiments are performed and the experimental results have demonstrated that the proposed model can significantly improve the accuracy and robustness for visual localization in dynamic environments.<;/p><;p><;i>Note to Practitioners<;/i>-This article was motivated by the visual localization problem in dynamic environments. Visual localization is well applied in many robotic fields such as path planning and exploration as the basic capability for a mobile robot. In the GPS-denied environments, one robot needs to localize itself through perceiving the unknown environment based on a visual sensor. In real-world scenes, the existence of the moving objects will significantly degrade the localization accuracy, which makes the robot implementation unreliable. In this article, an SMR model is designed to handle this problem. Once receiving a frame, the proposed model divides it into dynamic and static regions through a Bayesian framework. The dynamic regions are eliminated, while the static ones are maintained and fed into a feature-based visual SLAM system for further visual localization. The proposed method greatly improves the localization accuracy in dynamic environments and guarantees the robustness for robotic implementation.

Abstract. In this paper, we present a method for visual localization and pose estimation based on 3D image spaces. The method works in indoor and outdoor environments and does not require the presence of control points or markers. The method is evaluated with different sensors in an outdoor and an indoor test field. The results of our research show the viability of single image localization with absolute position accuracies at the decimetre level for outdoor environments and 5 cm or better for indoor environments. However, the evaluation also revealed a number of limitations of single image visual localization in real-world environments. Some of them could be addressed by an alternative AR-based localization approach, which we also present and compare in this paper. We then discuss the strengths and weaknesses of the two approaches and show possibilities for combining them to obtain accurate and robust visual localization in an absolute coordinate frame.

Robust visual localization in traffic scenes is a fundamental problem for self-driving vehicles. However, it is still challenging to achieve accurate localization performance because of drastic viewpoint and illumination changes. To address the issues, we design a novel monocular localization framework based on a light-weight prior map, called BSP-MonoLoc, which leverages the 2D semantic primitives from the monocular images and the 3D basic semantic primitives from the prior map. These primitives are commonly available but lack of distinctive signature. To effectively make associations between the 2D and 3D primitives and refine the vehicle’s pose, we adopt an iterative optimization method, where an efficient hierarchical sample strategy is designed to give a good initial prediction for the associations and the pose. Experimental results on the KAIST dataset and our dataset demonstrate the proposed method can achieve high localization accuracy and run at a real-time performance.

We present an illumination-robust visual localization algorithm for Astrobee, a free-flying robot designed to autonomously navigate on the International Space Station (ISS). Astrobee localizes with a monocular camera and a pre-built sparse map composed of natural visual features. Astrobee must perform tasks not only during the day, but also at night when the ISS lights are dimmed. However, the localization performance degrades when the observed lighting conditions differ from the conditions when the sparse map was built. We investigate and quantify the effect of lighting variations on visual feature-based localization systems, and discover that maps built in darker conditions can also be effective in bright conditions, but the reverse is not true. We extend Astrobee's localization algorithm to make it more robust to changing-light environments on the ISS by automatically recognizing the current illumination level, and selecting an appropriate map and camera exposure time. We extensively evaluate the proposed algorithm through experiments on Astrobee.

Camera localization from monocular images has been a long-standing problem, but its robustness in dynamic environments is still not adequately addressed. Compared with classic geometric approaches, modern CNN-based methods (e.g. PoseNet) have manifested the reliability against illumination or viewpoint variations, but they still have the following limitations. First, foreground moving objects are not explicitly handled, which results in poor performance and instability in dynamic environments. Second, the output for each image is a point estimate without uncertainty quantification. In this paper, we propose a framework which can be generally applied to existing CNN-based pose regressors to improve their robustness in dynamic environments. The key idea is a prior guided dropout module coupled with a self-attention module which can guide CNNs to ignore foreground objects during both training and inference. Additionally, the dropout module enables the pose regressor to output multiple hypotheses from which the uncertainty of pose estimates can be quantified and leveraged in the following uncertainty-aware pose-graph optimization to improve the robustness further. We achieve an average accuracy of 9.98m/3.63° on RobotCar dataset, which outperforms the state-of-the-art method by 62.97%/47.08%. The source code of our implementation is available at https://github.com/zju3dv/RVL-dynamic.

Simultaneous Localization and Mapping (SLAM) has developed as a fundamental method for intelligent robot perception over the past decades. Most of the existing feature-based SLAM systems relied on traditional hand-crafted visual features and a strong static world assumption, which makes these systems vulnerable in complex dynamic environments. In this paper, we propose a robust monocular SLAM system by combining geometry-based methods with two convolutional neural networks. Specifically, a lightweight deep local feature detection network is proposed as the system front-end, which can efficiently generate keypoints and binary descriptors robust against variations in illumination and viewpoint. Besides, we propose a motion segmentation and depth estimation network for simultaneously predicting pixel-wise motion object segmentation and depth map, so that our system can easily discard dynamic features and reconstruct 3D maps without dynamic objects. The comparison against state-of-the-art methods on publicly available datasets shows the effectiveness of our system in highly dynamic environments.

An improved lightweight deep neural network with knowledge distillation for local feature extraction and visual localization using images and LiDAR point clouds

Visual localization plays a key role in various robot perception systems. Robust visual localization relies on reliable and repeatable local features to establish high quality point correspondences among images. This letter focuses on addressing two limitations of joint learning detector and descriptor. First, existing methods use independent structures and loss functions for keypoint detection and description separately, which poses difficulty in detecting keypoints corresponding to discriminative descriptors. Second, triplet samples are treated equally in most existing approaches, which limits the learning algorithm to obtain highly discriminative descriptors. In this letter, we propose Task-aligned SuperPoint (TaSP) to mitigate the above problems. First, we explicitly align descriptor and detector learning to improve the probability of being detected for those distinctive points. Second, we introduce a dynamic importance weighting module that calculates the weight of each triplet sample based on intrinsic and empirical importance, so as to make the network focus on the most informative triplets during the whole training process. In addition, we resort to 3D space to seek negative samples when forming triplets, which avoids the risk of selecting negatives from repetitive structures. State-of-the-art results on a variety of visual localization benchmarks demonstrate the superiority of our method.

Image classification is getting more attention in the area of computer vision. During the past few years, a lot of research has been done on image classification using classical machine learning and deep learning techniques. Presently, deep learning-based techniques have given stupendous results. The performance of a classification system depends on the quality of features extracted from an image. The better is the quality of extracted features, the more the accuracy will be. Although, numerous deep learning-based methods have shown enormous performance in image classification, still due to various challenges deep learning methods are not able to extract all the important information from the image. This results in a reduction in overall classification accuracy. The goal of the present research is to improve the image classification performance by combining the deep features extracted using popular deep convolutional neural network, VGG19, and various handcrafted feature extraction methods, i.e., SIFT, SURF, ORB, and Shi-Tomasi corner detector algorithm. Further, the extracted features from these methods are classified using various machine learning classification methods, i.e., Gaussian Naïve Bayes, Decision Tree, Random Forest, and eXtreme Gradient Boosting (XGBClassifier) classifier. The experiment is carried out on a benchmark dataset Caltech-101. The experimental results indicate that Random Forest using the combined features give 93.73% accuracy and outperforms other classifiers and methods proposed by other authors. The paper concludes that a single feature extractor whether shallow or deep is not enough to achieve satisfactory results. So, a combined approach using deep learning features and traditional handcrafted features is better for image classification.

Gait recognition provides a more convenient way for human identification, as it can identify person with less cooperation and intrusion compared to other biometric features. Current gait recognition frameworks utilize a template to extract temporal feature or regard the whole person as a unit, and they obtain limited temporal information and fine-grained features. To overcome this problem, we propose a network consisting of two parts: Temporal Feature Fusion (TFF) and Fine-grained Feature Extraction (FFE). First, we extract the most representative temporal information from raw gait sequences by TFF. Next, we use the idea of partial features on fused temporal features to extract more fine-grained spatial block features. It is worth mentioning that the proposed algorithm provides an effective feature extraction framework for complex gait recognition, as it focuses on the temporal fusion for representative information, and the extraction of the fine-grained spatial features. Extensive experiments illustrated that we have an outstanding performance on CASIA-B and mini-OUMVLP compared to other state-of-the-art methods including GaitSet and GaitNet. In particularly, the average rank-1 accuracy of all probe views on normal walking condition (NM) achieve 95.7%.

This thesis explores the problem of visual place recognition and localization for a mobile robot, particularly dealing with the challenges of simultaneous variations in scene appearance and camera viewpoint. The proposed methods draw inspiration from humans and make use of semantic cues to represent places. This approach enables effective place recognition from similar or opposing viewpoints, despite variations in scene appearance caused by different times of day or seasons. The research contributions presented in the thesis advance visual place recognition techniques, making them more useful for deployment in a wide range of robotic and autonomous vehicle scenarios.

Visual localization is an attractive problem that estimates the camera localization from database images based on the query image. It is a crucial task for various applications, such as autonomous vehicles, assistive navigation and augmented reality. The challenging issues of the task lie in various appearance variations between query and database images, including illumination variations, dynamic object variations and viewpoint variations. In order to tackle those challenges, Panoramic Annular Localizer into which panoramic annular lens and robust deep image descriptors are incorporated is proposed in this paper. The panoramic annular images captured by the single camera are processed and fed into the NetVLAD network to form the active deep descriptor, and sequential matching is utilized to generate the localization result. The experiments carried on the public datasets and in the field illustrate the validation of the proposed system.

SADGFeat: Learning local features with layer spatial attention and domain generalization

The real-world deployment of fully autonomous mobile robots depends on a robust simultaneous localization and mapping (SLAM) system, capable of handling dynamic environments, where objects are moving in front of the robot, and changing environments, where objects are moved or replaced after the robot has already mapped the scene. This paper proposes Changing-SLAM, a method for robust Visual SLAM in both dynamic and changing environments. This is achieved by using a Bayesian filter combined with a long-term data association algorithm. Also, it employs an efficient algorithm for dynamic keypoints filtering based on object detection that correctly identifies features inside the bounding box that are not dynamic, preventing a depletion of features that could cause lost tracks. Furthermore, a new dataset was developed with RGB-D data specially designed for the evaluation of changing environments on an object level, called PUC-USP dataset. Six sequences were created using a mobile robot, an RGB-D camera and a motion capture system. The sequences were designed to capture different scenarios that could lead to a tracking failure or map corruption. Changing-SLAM does not assume a given camera pose or a known map, being also able to operate in real time. The proposed method was evaluated using benchmark datasets and compared with other state-of-the-art methods, proving to be highly accurate.

This paper describes an approach for lip tracking using dynamic deformable templates. The objective is to track lip parameters from an audio-visual corpus recording a voice talent who is reading text prompts in a natural and expressive way. The corpus presents challenges to the conventional method of lip tracking with deformable templates. This is because natural and expressive speech includes relatively large motions of the head and the lips. The head motions lead to changes in the illumination of the face region and changes in the observed lip shape. In addition, emphatic pronunciations lead to large changes in the lip shape. Video frames that are affected by face illumination changes present additional difficulty in locating the mouth region (i.e. region of interest, ROI). Video frames that are affected by changes in lip shapes present additional deviations from the lip templates and hence lower tracking accuracies. Our proposed method incorporates dynamicity in the deformable templates to render them adaptive to changes in head pose, face illumination and lip shapes. Experiments show that dynamic deformable templates consistently outperform the conventional deformable templates in lip tracking.