Raw Depth Images Research Articles

NGHIÊN CỨU XÂY DỰNG MÔ HÌNH NHẬN DIỆN VẬT THỂ TRONG SUỐT DỰA TRÊN THỊ GIÁC LẬP THỂ VÀ TRÍ TUỆ NHÂN TẠO

The article focuses on researching the construction of a transparent object (glass) recognition model based on the application of computer vision techniques and artificial intelligence models. Stereo Matching image processing techniques have been used to build a raw depth image from a Stereo Camera. The goal is to reconstruct the depth image, recover in-depth information, and generate a complete depth image to effectively identify the position of transparent objects in reality. Additionally, the research involves designing a software interface for observing depth images, point clouds, and controlling the robotic arm for object grasping in threedimensional space. The following results were obtained: The quality of the depth image reconstruction model is improved compared to the ClearGrasp model when evaluated on ClearGrasp datasets; Determine orientations on how to improve models and algorithms for reconstructing depth images in a more quantitative manner. The success rate of picking up a glass cup is over 90% in cases of objects on the floor; This rate reaches over 70% when objects are placed at different heights. The software interface displays detailed information and facilitates communication, controlling depth images, point clouds, and position graphs (x, y, z). At the same time, it is easy to interact and convenient during the experiment

RDFC-GAN: RGB-Depth Fusion CycleGAN for Indoor Depth Completion.

Raw depth images captured in indoor scenarios frequently exhibit extensive missing values due to the inherent limitations of the sensors and environments. For example, transparent materials frequently elude detection by depth sensors; surfaces may introduce measurement inaccuracies due to their polished textures, extended distances, and oblique incidence angles from the sensor. The presence of incomplete depth maps imposes significant challenges for subsequent vision applications, prompting the development of numerous depth completion techniques to mitigate this problem. Numerous methods excel at reconstructing dense depth maps from sparse samples, but they often falter when faced with extensive contiguous regions of missing depth values, a prevalent and critical challenge in indoor environments. To overcome these challenges, we design a novel two-branch end-to-end fusion network named RDFC-GAN, which takes a pair of RGB and incomplete depth images as input to predict a dense and completed depth map. The first branch employs an encoder-decoder structure, by adhering to the Manhattan world assumption and utilizing normal maps from RGB-D information as guidance, to regress the local dense depth values from the raw depth map. The other branch applies an RGB-depth fusion CycleGAN, adept at translating RGB imagery into detailed, textured depth maps while ensuring high fidelity through cycle consistency. We fuse the two branches via adaptive fusion modules named W-AdaIN and train the model with the help of pseudo depth maps. Comprehensive evaluations on NYU-Depth V2 and SUN RGB-D datasets show that our method significantly enhances depth completion performance particularly in realistic indoor settings.

Autonomous Robotic Bin Picking Platform Generated From Human Demonstration and YOLOv5

Abstract Vision-based robots have been utilized for pick-and-place operations by their ability to find object poses. As they progress into handling a variety of objects with cluttered state, more flexible and lightweight operations have been presented. In this paper, an autonomous robotic bin-picking platform is proposed. It combines human demonstration with a collaborative robot for the flexibility of the objects and YOLOv5 neural network model for faster object localization without prior computer-aided design models or dataset in the training. After a simple human demonstration of which target object to pick and place, the raw color and depth images were refined, and the one on top of the bin was utilized to create synthetic images and annotations for the YOLOv5 model. To pick up the target object, the point cloud was lifted using the depth data corresponding to the result of the trained YOLOv5 model, and the object pose was estimated by matching them with Iterative Closest Points (ICP) algorithm. After picking up the target object, the robot placed it where the user defined it in the previous human demonstration stage. From the result of experiments with four types of objects and four human demonstrations, it took a total of 0.5 s to recognize the target object and estimate the object pose. The success rate of object detection was 95.6%, and the pick-and-place motion of all the found objects was successful.

A Depth-Based Hybrid Approach for Safe Flight Corridor Generation in Memoryless Planning.

This paper presents a depth-based hybrid method to generate safe flight corridors for a memoryless local navigation planner. It is first proposed to use raw depth images as inputs in the learning-based object-detection engine with no requirement for map fusion. We then employ an object-detection network to directly predict the base of polyhedral safe corridors in a new raw depth image. Furthermore, we apply a verification procedure to eliminate any false predictions so that the resulting collision-free corridors are guaranteed. More importantly, the proposed mechanism helps produce separate safe corridors with minimal overlap that are suitable to be used as space boundaries for path planning. The average intersection of union (IoU) of corridors obtained by the proposed algorithm is less than 2%. To evaluate the effectiveness of our method, we incorporated it into a memoryless planner with a straight-line path-planning algorithm. We then tested the entire system in both synthetic and real-world obstacle-dense environments. The obtained results with very high success rates demonstrate that the proposed approach is highly capable of producing safe corridors for memoryless local planning.

Optimal Transport for Unsupervised Denoising Learning.

Recently, much progress has been made in unsupervised denoising learning. However, existing methods more or less rely on some assumptions on the signal and/or degradation model, which limits their practical performance. How to construct an optimal criterion for unsupervised denoising learning without any prior knowledge on the degradation model is still an open question. Toward answering this question, this work proposes a criterion for unsupervised denoising learning based on the optimal transport theory. This criterion has favorable properties, e.g., approximately maximal preservation of the information of the signal, whilst achieving perceptual reconstruction. Furthermore, though a relaxed unconstrained formulation is used in practical implementation, we prove that the relaxed formulation in theory has the same solution as the original constrained formulation. Experiments on synthetic and real-world data, including realistic photographic, microscopy, depth, and raw depth images, demonstrate that the proposed method even compares favorably with supervised methods, e.g., approaching the PSNR of supervised methods while having better perceptual quality. Particularly, for spatially correlated noise and realistic microscopy images, the proposed method not only achieves better perceptual quality but also has higher PSNR than supervised methods. Besides, it shows remarkable superiority in harsh practical conditions with complex noise, e.g., raw depth images. Code is available at https://github.com/wangweiSJTU/OTUR.

An Open-Source Low-Cost Mobile Robot System With an RGB-D Camera and Efficient Real-Time Navigation Algorithm

Currently, mobile robots are developing rapidly and are finding numerous applications in the industry. However, several problems remain related to their practical use, such as the need for expensive hardware and high power consumption levels. In this study, we build a low-cost indoor mobile robot platform that does not include a LiDAR or a GPU. Then, we design an autonomous navigation architecture that guarantees real-time performance on our platform with an RGB-D camera and a low-end off-the-shelf single board computer. The overall system includes SLAM, global path planning, ground segmentation, and motion planning. The proposed ground segmentation approach extracts a traversability map from raw depth images for the safe driving of low-body mobile robots. We apply both rule-based and learning-based navigation policies using the traversability map. Running sensor data processing and other autonomous driving components simultaneously, our navigation policies perform rapidly at a refresh rate of 18 Hz for control command, whereas other systems have slower refresh rates. Our methods show better performances than current state-of-the-art navigation approaches within limited computation resources as shown in 3D simulation tests. In addition, we demonstrate the applicability of our mobile robot system through successful autonomous driving in an indoor environment.

Multiple Cylinder Extraction from Organized Point Clouds.

Most man-made objects are composed of a few basic geometric primitives (GPs) such as spheres, cylinders, planes, ellipsoids, or cones. Thus, the object recognition problem can be considered as one of geometric primitives extraction. Among the different geometric primitives, cylinders are the most frequently used GPs in real-world scenes. Therefore, cylinder detection and extraction are of great importance in 3D computer vision. Despite the rapid progress of cylinder detection algorithms, there are still two open problems in this area. First, a robust strategy is needed for the initial sample selection component of the cylinder extraction module. Second, detecting multiple cylinders simultaneously has not yet been investigated in depth. In this paper, a robust solution is provided to address these problems. The proposed solution is divided into three sub-modules. The first sub-module is a fast and accurate normal vector estimation algorithm from raw depth images. With the estimation method, a closed-form solution is provided for computing the normal vector at each point. The second sub-module benefits from the maximally stable extremal regions (MSER) feature detector to simultaneously detect cylinders present in the scene. Finally, the detected cylinders are extracted using the proposed cylinder extraction algorithm. Quantitative and qualitative results show that the proposed algorithm outperforms the baseline algorithms in each of the following areas: normal estimation, cylinder detection, and cylinder extraction.

Robot obstacle avoidance system using deep reinforcement learning

PurposeMost manufacturing plants choose the easy way of completely separating human operators from robots to prevent accidents, but as a result, it dramatically affects the overall quality and speed that is expected from human–robot collaboration. It is not an easy task to ensure human safety when he/she has entered a robot’s workspace, and the unstructured nature of those working environments makes it even harder. The purpose of this paper is to propose a real-time robot collision avoidance method to alleviate this problem.Design/methodology/approachIn this paper, a model is trained to learn the direct control commands from the raw depth images through self-supervised reinforcement learning algorithm. To reduce the effect of sample inefficiency and safety during initial training, a virtual reality platform is used to simulate a natural working environment and generate obstacle avoidance data for training. To ensure a smooth transfer to a real robot, the automatic domain randomization technique is used to generate randomly distributed environmental parameters through the obstacle avoidance simulation of virtual robots in the virtual environment, contributing to better performance in the natural environment.FindingsThe method has been tested in both simulations with a real UR3 robot for several practical applications. The results of this paper indicate that the proposed approach can effectively make the robot safety-aware and learn how to divert its trajectory to avoid accidents with humans within the workspace.Research limitations/implicationsThe method has been tested in both simulations with a real UR3 robot in several practical applications. The results indicate that the proposed approach can effectively make the robot be aware of safety and learn how to change its trajectory to avoid accidents with persons within the workspace.Originality/valueThis paper provides a novel collision avoidance framework that allows robots to work alongside human operators in unstructured and complex environments. The method uses end-to-end policy training to directly extract the optimal path from the visual inputs for the scene.

Learn to Navigate Autonomously Through Deep Reinforcement Learning

In this article, we propose a deep reinforcement learning (DRL) algorithm as well as a novel tailored neural network architecture for mobile robots to learn navigation policies autonomously. We first introduce a new feature extractor to better acquire critical spatiotemporal features from raw depth images. In addition, we present a double-source scheme so that the experiences are collected from both the proposed model and a conventional planner alternatively based on a switching criterion to provide more diverse and comprehensive samples for learning. Moreover, we also propose a dual-soft-actor–critic architecture to train two sets of networks with different purposes simultaneously. Specifically, the primary network aims to learn the autonomous navigation policy, while the secondary network aims to learn the depth feature extractor. In this way, the learning performance can be improved through decoupling representation learning from policy learning and training the feature extractor separately with more specific goals. Experimental results have demonstrated the remarkable performance of the proposed model. It outperforms both the conventional and the state-of-the-art DRL-based methods consistently in terms of success rate. Meanwhile, it also exhibits higher trajectory quality and better generalization capability compared to existing DRL-based methods. Videos of our experiments are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://youtu.be/evjU6bOU3UY</uri> or OneDrive.

Volumetric Occupancy Mapping With Probabilistic Depth Completion for Robotic Navigation

In robotic applications, a key requirement for safe and efficient motion planning is the ability to map obstacle-free space in unknown, cluttered 3D environments. However, commodity-grade RGB-D cameras commonly used for sensing fail to register valid depth values on shiny, glossy, bright, or distant surfaces, leading to missing data in the map. To address this issue, we propose a framework leveraging probabilistic depth completion as an additional input for spatial mapping. We introduce a deep learning architecture providing uncertainty estimates for the depth completion of RGB-D images. Our pipeline exploits the inferred missing depth values and depth uncertainty to complement raw depth images and improve the speed and quality of free space mapping. Evaluations on synthetic data show that our approach maps significantly more correct free space with relatively low error when compared against using raw data alone in different indoor environments; thereby producing more complete maps that can be directly used for robotic navigation tasks. The performance of our framework is validated using real-world data.

Practical aspects of detection and grasping objects by a mobile manipulating robot

PurposeThe purpose of this paper is to propose and evaluate the method for grasping a defined set of objects in an unstructured environment. To this end, the authors propose the method of integrating convolutional neural network (CNN)-based object detection and the category-free grasping method. The considered scenario is related to mobile manipulating platforms that move freely between workstations and manipulate defined objects. In this application, the robot is not positioned with respect to the table and manipulated objects. The robot detects objects in the environment and uses grasping methods to determine the reference pose of the gripper.Design/methodology/approachThe authors implemented the whole pipeline which includes object detection, grasp planning and motion execution on the real robot. The selected grasping method uses raw depth images to find the configuration of the gripper. The authors compared the proposed approach with a representative grasping method that uses a 3D point cloud as an input to determine the grasp for the robotic arm equipped with a two-fingered gripper. To measure and compare the efficiency of these methods, the authors measured the success rate in various scenarios. Additionally, they evaluated the accuracy of object detection and pose estimation modules.FindingsThe performed experiments revealed that the CNN-based object detection and the category-free grasping methods can be integrated to obtain the system which allows grasping defined objects in the unstructured environment. The authors also identified the specific limitations of neural-based and point cloud-based methods. They show how the determined properties influence the performance of the whole system.Research limitations/implicationsThe authors identified the limitations of the proposed methods and the improvements are envisioned as part of future research.Practical implicationsThe evaluation of the grasping and object detection methods on the mobile manipulating robot may be useful for all researchers working on the autonomy of similar platforms in various applications.Social implicationsThe proposed method increases the autonomy of robots in applications in the small industry which is related to repetitive tasks in a noisy and potentially risky environment. This allows reducing the human workload in these types of environments.Originality/valueThe main contribution of this research is the integration of the state-of-the-art methods for grasping objects with object detection methods and evaluation of the whole system on the industrial robot. Moreover, the properties of each subsystem are identified and measured.

A multi-cue guidance network for depth completion

Depth completion is to predict a dense depth image from a raw sparse depth image with missing values, which is an important yet challenging problem in a myriad of vision, robotics, and multimedia applications. While previous studies have made substantial progress in this issue, most of them directly fuse RGB and depth features without considering the multiple latent cues in data. In this paper, we propose a multi-cue guidance network model for depth completion, which introduces multi-cue features to guide the regression of residual values. This network includes two major parts: multi-cue guidance structure and residual regression structure. The multi-cue guidance structure composed of two parallel convolutional sampling streams extracts features from the raw depth image and the corresponding RGB image, respectively. The extracted features are used as multi-cue guidance to combine with the depth image to feed the residual regression structure, where the estimated residual values are combined with the raw depth values to output the final dense depth values. By virtue of the multi-cue guidance and residual regression, the proposed model can leverage the multiple latent cues in the data to predict more accurate depth values. The proposed method was tested on two challenging datasets of outdoor and indoor scenes: KITTI benchmark and NYUv2 dataset. The experimental results show that the proposed method outperforms the comparison methods. Extensive ablation studies demonstrate the effectiveness of different modules in the proposed method.

Reconstruction-based 6D pose estimation for robotic assembly.

Pose estimation is important for many robotic applications including bin picking and robotic assembly and collaboration. However, robust and accurate estimation of the poses of industrial objects is a challenging task owing to the various object shapes and complex working environments. This paper presents a method of estimating the poses of narrow and elongated industrial objects with a low-cost RGB-D (depth and color) camera to guide the process of robotic assembly. The proposed method comprises three main steps: reconstruction involved in preprocessing, pose initialization with geometric features, and tracking aided by contour cues. Pose tracking is coupled with real-time dense reconstruction, which can synthesize a smooth depth image as a substitute for the raw depth image. Because industrial objects (e.g., fork and adapter) feature mostly planar structures, primitive geometric features, such as three-dimensional planes, are extracted from the point cloud and utilized to induce a promising initial pose. For robust tracking of the adapter consisting of narrow and elongated planes, the dense surface correspondences are combined with sparse contour correspondences in the refinement scheme. This combination allows for a satisfactory tolerance to the initial guess in the pose tracking phase. The experimental results demonstrate the feasibility of the proposed method.

Combining learned and analytical models for predicting action effects from sensory data

One of the most basic skills a robot should possess is predicting the effect of physical interactions with objects in the environment. This enables optimal action selection to reach a certain goal state. Traditionally, dynamics are approximated by physics-based analytical models. These models rely on specific state representations that may be hard to obtain from raw sensory data, especially if no knowledge of the object shape is assumed. More recently, we have seen learning approaches that can predict the effect of complex physical interactions directly from sensory input. It is, however, an open question how far these models generalize beyond their training data. In this work, we investigate the advantages and limitations of neural-network-based learning approaches for predicting the effects of actions based on sensory input and show how analytical and learned models can be combined to leverage the best of both worlds. As physical interaction task, we use planar pushing, for which there exists a well-known analytical model and a large real-world dataset. We propose the use of a convolutional neural network to convert raw depth images or organized point clouds into a suitable representation for the analytical model and compare this approach with using neural networks for both, perception and prediction. A systematic evaluation of the proposed approach on a very large real-world dataset shows two main advantages of the hybrid architecture. Compared with a pure neural network, it significantly (i) reduces required training data and (ii) improves generalization to novel physical interaction.

A Comparative Study in Real-Time Scene Sonification for Visually Impaired People.

In recent years, with the development of depth cameras and scene detection algorithms, a wide variety of electronic travel aids for visually impaired people have been proposed. However, it is still challenging to convey scene information to visually impaired people efficiently. In this paper, we propose three different auditory-based interaction methods, i.e., depth image sonification, obstacle sonification as well as path sonification, which convey raw depth images, obstacle information and path information respectively to visually impaired people. Three sonification methods are compared comprehensively through a field experiment attended by twelve visually impaired participants. The results show that the sonification of high-level scene information, such as the direction of pathway, is easier to learn and adapt, and is more suitable for point-to-point navigation. In contrast, through the sonification of low-level scene information, such as raw depth images, visually impaired people can understand the surrounding environment more comprehensively. Furthermore, there is no interaction method that is best suited for all participants in the experiment, and visually impaired individuals need a period of time to find the most suitable interaction method. Our findings highlight the features and the differences of three scene detection algorithms and the corresponding sonification methods. The results provide insights into the design of electronic travel aids, and the conclusions can also be applied in other fields, such as the sound feedback of virtual reality applications.

SVM-based waist circumference estimation using Kinect

Background and objectiveConventional anthropometric studies using Kinect depth sensors have concentrated on estimating the distances between two points such as height. This paper deals with a novel waist measurement method using SVM regression, further widening spectrum of Kinect's potential applications. Waist circumference is a key index for the diagnosis of abdominal obesity, which has been linked to metabolic syndromes and other related diseases. Yet, the existing measuring method, tape measure, requires a trained personnel and is therefore costly and time-consuming. MethodsA dataset was constructed by recording both 30 frames of Kinect depth image and careful tape measurement of 19 volunteers by a clinical investigator. This paper proposes a new SVM regressor-based approach for estimating waist circumference. A waist curve vector is extracted from a raw depth image using joint information provided by Kinect SDK. To avoid overfitting, a data augmentation technique is devised. The 30 frontal vectors and 30 backside vectors, each sampled for 1 s per person, are combined to form 900 waist curve vectors and a total of 17,100 samples were collected from 19 individuals. On an individual basis, we performed leave-one-out validation using the SVM regressor with the tape measurement-gold standard of waist circumference measurement-values labeled as ground-truth. On an individual basis, we performed leave-one-out validation using the SVM regressor with the tape measurement-gold standard of waist circumference measurement-values labeled as ground-truth. ResultsThe mean error of the SVM regressor was 4.62 cm, which was smaller than that of the geometric estimation method. Potential uses are discussed. ConclusionsA possible method for measuring waist circumference using a depth sensor is demonstrated through experimentation. Methods for improving accuracy in the future are presented. Combined with other potential applications of Kinect in healthcare setting, the proposed method will pave the way for patient-centric approach of delivering care without laying burdens on patients.

CLUBS: An RGB-D dataset with cluttered box scenes containing household objects

With the progress of machine learning, the demand for realistic data with high-quality annotations has been thriving. In order to generalize well, considerable amounts of data are required, especially realistic ground-truth data, for tasks such as object detection and scene segmentation. Such data can be difficult, time-consuming, and expensive to collect. This article presents a dataset of household objects and box scenes commonly found in warehouse environments. The dataset was obtained using a robotic setup with four different cameras. It contains reconstructed objects and scenes, as well as raw RGB and depth images, camera poses, pixel-wise labels of objects directly in the RGB images, and 3D bounding boxes with poses in the world frame. Furthermore, raw calibration data are provided, together with the intrinsic and extrinsic parameters for all the sensors. By providing object labels as pixel-wise masks, 3D, and 2D object bounding boxes, this dataset is useful for both object recognition and instance segmentation. The realistic scenes provided will serve for learning-based algorithms applied to scenarios where boxes of objects are often found, such as in the logistics sector. Both the dataset and the tools for data processing are published and available online.

High-Quality 3D Reconstruction With Depth Super-Resolution and Completion

The 3D reconstruction is an important topic in computer vision with many applications, such as robotics and augmented reality. Since the raw depth images captured by consumer RGB-D cameras are often low resolution (LR), noisy, and incomplete. How to obtain high-quality 3D models with a consumer RGB-D camera is still a challenge for the existing systems. In this paper, we propose a new depth super-resolution and completion method implemented in a deep learning framework and build a high-quality 3D reconstruction system. We first improve the resolution of LR depth image with a depth super-resolution network and remove the outliers in high-resolution (HR) depth image based on gradient saliency. To further enhance the quality of HR depth image with the guide of HR color image, we learn surface normal and occlusion boundary images from the corresponding HR color image through two deep fully convolutional networks. In particular, the blurriness of HR color image is also detected and pixel-wise quantized. Finally, we obtain a completed HR depth image by optimizing the HR depth image with the surface normal, occlusion boundary, and color image blurriness. We have carried out qualitative and quantitative evaluations with baseline methods on public datasets. The experimental results demonstrate that our method has better performance both on single depth image enhancement and 3D reconstruction.

Learn to Steer through Deep Reinforcement Learning.

It is crucial for robots to autonomously steer in complex environments safely without colliding with any obstacles. Compared to conventional methods, deep reinforcement learning-based methods are able to learn from past experiences automatically and enhance the generalization capability to cope with unseen circumstances. Therefore, we propose an end-to-end deep reinforcement learning algorithm in this paper to improve the performance of autonomous steering in complex environments. By embedding a branching noisy dueling architecture, the proposed model is capable of deriving steering commands directly from raw depth images with high efficiency. Specifically, our learning-based approach extracts the feature representation from depth inputs through convolutional neural networks and maps it to both linear and angular velocity commands simultaneously through different streams of the network. Moreover, the training framework is also meticulously designed to improve the learning efficiency and effectiveness. It is worth noting that the developed system is readily transferable from virtual training scenarios to real-world deployment without any fine-tuning by utilizing depth images. The proposed method is evaluated and compared with a series of baseline methods in various virtual environments. Experimental results demonstrate the superiority of the proposed model in terms of average reward, learning efficiency, success rate as well as computational time. Moreover, a variety of real-world experiments are also conducted which reveal the high adaptability of our model to both static and dynamic obstacle-cluttered environments.

An Occlusion-Aware Framework for Real-Time 3D Pose Tracking.

Random forest-based methods for 3D temporal tracking over an image sequence have gained increasing prominence in recent years. They do not require object’s texture and only use the raw depth images and previous pose as input, which makes them especially suitable for textureless objects. These methods learn a built-in occlusion handling from predetermined occlusion patterns, which are not always able to model the real case. Besides, the input of random forest is mixed with more and more outliers as the occlusion deepens. In this paper, we propose an occlusion-aware framework capable of real-time and robust 3D pose tracking from RGB-D images. To this end, the proposed framework is anchored in the random forest-based learning strategy, referred to as RFtracker. We aim to enhance its performance from two aspects: integrated local refinement of random forest on one side, and online rendering based occlusion handling on the other. In order to eliminate the inconsistency between learning and prediction of RFtracker, a local refinement step is embedded to guide random forest towards the optimal regression. Furthermore, we present an online rendering-based occlusion handling to improve the robustness against dynamic occlusion. Meanwhile, a lightweight convolutional neural network-based motion-compensated (CMC) module is designed to cope with fast motion and inevitable physical delay caused by imaging frequency and data transmission. Finally, experiments show that our proposed framework can cope better with heavily-occluded scenes than RFtracker and preserve the real-time performance.