Robotic grasping recognition using multi-modal deep extreme learning machine

Recognizing which part of an object is graspable or not is important for intelligent robot to perform some complicated tasks. In order to obtain good grasping performance, learning rich representations efficiently from multi-modal RGB-D images is crucial. To address this problem, in this paper, we propose an effective multi-modal deep extreme learning machine structure. In this structure, unsupervised hierarchical extreme learning machine (ELM) is conducted for feature extraction for RGB and depth modalities separately. Then, the shared layer is developed by combining both RGB and depth features. Finally, the ELM is used as supervised feature classifier for final decision. Experimental validation on Cornell grasping dataset illustrates that the proposed multiple modality fusion method achieves better grasp recognition performance.     

复杂场景点云数据的6D位姿估计深度学习网络

针对工业上常见的弱纹理、散乱摆放复杂场景下点云目标机器人抓取问题,该文提出一种6D位姿估计深度学习网络。首先,模拟复杂场景下点云目标多姿态随机摆放的物理环境,生成带真实标签的数据集;进而,设计了6D位姿估计深度学习网络模型,提出多尺度点云分割网络(MPCS-Net),直接在完整几何点云上进行点云实例分割,解决了对RGB信息和点云分割预处理的依赖问题。然后,提出多层特征姿态估计网(MFPE-Net),有效地解决了对称物体的位姿估计问题。最后,实验结果和分析证实了,相比于传统的点云配准方法和现有的切分点云的深度学习位姿估计方法,所提方法取得了更高的准确率和更稳定性能,并且在估计对称物体位姿时有较强的鲁棒性。     

An experimental approach to robotic grasping using a connectionistarchitecture and generic grasping functions

An experimental approach to robotic grasping is presented. This approach is based on developing a generic representation of grasping rules, which allows learning them from experiments between the object and the robot. A modular connectionist design arranged in subsumption layers is used to provide a mapping between sensory inputs and robot actions. Reinforcement feedback is used to select between different grasping rules and to reduce the number of failed experiments. This is particularly critical for applications in the personal service robot environment. Simulated experiments on a 15-object database show that the system is capable of learning grasping rules for each object in a finite number of experiments as well as generalizing from experiments on one object to grasping from another     

Visual servoing on unknown objects

We study visual servoing in a framework of detection and grasping of unknown objects. Classically, visual servoing has been used for applications where the object to be servoed on is known to the robot prior to the task execution. In addition, most of the methods concentrate on aligning the robot hand with the object without grasping it. In our work, visual servoing techniques are used as building blocks in a system capable of detecting and grasping unknown objects in natural scenes. We show how different visual servoing techniques facilitate a complete grasping cycle.     

面向移动端的单阶段端到端目标检测压缩算法

近30年间,深度学习异军突起。它在各项计算机视觉任务中都取得了令人瞩目的进步,加之大量高质多样化数据的出现,使得各种依赖数据的目标检测方法重现曙光。然而,这些深度网络算法通常需要大量数据来支持数百亿参数的计算,其运行效率较低并且对存储空间的要求越来越高,使得在小型设备或移动端中无法嵌入大型神经网络。因此,本文提出优化目标检测算法以适应移动端环境,利用CNN卷积核多样性和可分离的原理,应用深度可分离卷积（Depthwise Separable Convolution）结构的理论,提出单阶段-端到端目标检测压缩网络DW-YOLOv3。最后,在带有详细标注的地面观测实况大规模基准数据集VisDrone2018数据集上的结果表明,本文提出的改进单阶段-可分离卷积目标检测压缩网络算法可以将网络参数压缩8-9倍,由于其增加了整体网络的深度,在对网络整体性能影响较小的同时提升了对无人机视角图像中小目标物体的识别性能。      

A Mixed Mode Neural Network Circuitry for Object Recognition Application

A general purpose Conic Section Function Neural Network (CSFNN) circuitry in Very Large Scale Integration (VLSI) has been designed for an object recognition application. CSFNN is capable of making open and closed decision regions by combining the propagation rules of Radial Basis Functions (RBF) and Multilayer Perceptrons (MLP) on a single neural network with a unique propagation rule. Chip-in-the-loop learning technique was used during the training process. Utilizing mixed-mode hardware techniques, the inputs of the network and the feedforward signals are all analog while the control unit and storage of the network parameters are fully digital. CSFNN circuitry architecture is problem independent and consists of 16 inputs, 16 hidden layer neurons and 8 outputs. Inheriting the merits of CSFNN, the circuitry has good recognition performance on several objects with invariance to pose, lighting, and brightness. The designed hardware achieved a good recognition performance by means of both accuracy and computational time comparable to CSFNN software.     

Modulation Recognition in Maritime Multipath Channels: A Blind Equalization-Aided Deep Learning Approach

Modulation recognition has been long investigated in the literature,however,the performance could be severely degraded in multipath fading channels especially for high-order Quadrature Amplitude Modulation(QAM)signals.This could be a critical problem in the broadband maritime wireless communications,where various propagation paths with large differences in the time of arrival are very likely to exist.Specifically,multiple paths may stem from the direct path,the reflection paths from the rough sea surface,and the refraction paths from the atmospheric duct,respectively.To address this issue,we propose a novel blind equalization-aided deep learning(DL)approach to recognize QAM signals in the presence of multipath propagation.The proposed approach consists of two modules:A blind equalization module and a subsequent DL network which employs the structure of ResNet.With predefined searching step-sizes for the blind equalization algorithm,which are designed according to the set of modulation formats of interest,the DL network is trained and tested over various multipath channel parameter settings.It is shown that as compared to the conventional DL approaches without equalization,the proposed method can achieve an improvement in the recognition accuracy up to 30%in severe multipath scenarios,especially in the high SNR regime.Moreover,it efficiently reduces the number of training data that is required.     

基于卷积神经网络迁移学习的飞机目标识别

飞机目标识别是地面情报系统的一项重要关键技术。近年来火热的深度学习方法,如卷积神经网络,展现出对于图像识别任务的优越性能。但是,训练卷积神经网络需要大量的带标签样本以估计规模庞大的模型参数,因而限制了其在雷达目标识别领域中的应用。针对飞机目标识别中的小样本问题,文中引入适用于有限数据场景的迁移学习技术,预先在其他大样本高分辨距离像数据上训练一个初始卷积神经网络模型,再结合当前飞机目标识别任务调优模型参数。在实测数据上的实验结果显示,与仅使用卷积神经网络的方法相比,所提方法可显著提升识别准确率,验证了方法的有效性。     

采用深度卷积神经网络方法的毫米波图像目标检测

通过收集大量的毫米波图像并建立相应的人体数据集进行检测,提出基于Faster R-CNN深度学习的方法检测隐藏于人体上的危险物品。该方法将区域建议网络和VGG19训练卷积神经网络模型相结合,构建了面向毫米波图像目标检测的深度卷积神经网络。为了提高毫米波图像的处理能力,采用Caffe深度学习框架在图形处理单元上进行训练和测试。实验结果证明了基于Faster R-CNN深度卷积神经网络的目标检测方法能有效检测毫米波图像中的危险物品,并且目标检测的平均准确率约94%,检测速度约为6 frame/s,对毫米波安检系统的智能化发展有着极其重要的参考价值。     

Stochastic re-grasp planning for vision aided capture of deforming and moving object

In the last two decades several researchers have studied the problem of grasping of a moving rigid object based on vision data. However the problem of grasping a moving and deforming object still remains unsolved. In this paper we present the development of a fast algorithm for the computation of the optimal force closure grasp points on a slowly moving and deforming object. The main focus is to find the best grasp points as the object deforms, track its position at a future instant and then transfer grasp at that location. At first the potential grasping configurations satisfying force closure are evaluated through an objective function that maximizes the grasping span while minimizing the distance between the object centroid and the intersection of the fingertip normal. A population based stochastic search strategy is adopted for computing the optimal configurations and re-localizing them as the shape undergoes translations, rotations and scaling. Experiments have been conducted to prove that the object can be tracked in real time and the optimal grasp points determined so that a three finger robot can capture it. This method works in real time so it has great potential for application in industries for grasping objects whose shapes are not clearly defined (e.g. cloth), deforming objects, or objects that are partially occluded.     

深度学习在雷达通信目标识别中的应用框架

深度学习是当前人工神经网络领域的研究热点,广泛应用于字符识别、图像识别和语音识别等应用中。雷达通信目标识别是通信对抗的前提和关键。文中分析了模板匹配法、DS证据理论等传统通信目标识别方法的在特征提取、模型表达方面的不足,对深度学习神经网络在通信目标识别中的应用进行了初步探讨,并提出了一种基于深度学习的通信目标识别框架。该框架和思路同样适用于雷达对抗目标识别等问题,可为深度学习在雷达目标识别领域的应用提供支撑。     

TWin support tensor machines for MCs detection

Tensor representation is useful to reduce the overfitting problem in vector-based learning algorithm in pattern recognition.This is mainly because the structure information of objects in pattern analysis is a reasonable constraint to reduce the number of unknown parameters used to model a classifier.In this paper,we generalize the vector-based learning algorithm TWin Support Vector Machine (TWSVM)to the tensor-based method TWin Support Tensor Machines(TWSTM),which accepts general tensors as input.To examine the effectiveness of TWSTM,we implement the TWSTM method for Microcalcification Clusters (MCs) detection.In the tensor subspace domain,the MCs detection procedure is formulated as a supervised learning and classification problem.and TWSTM is used as a classifier to make decision for the presence of MCs or not.A large number of experiments were carried out to evaluate and compare the performance of the proposed MCs detection algorithm.By comparison with TWSVM,the tensor version reduces the overfitting problem.     

Transferring human grasping synergies to a robot

In this paper, a system for transferring human grasping skills to a robot is presented. In order to reduce the dimensionality of the grasp postures, we extracted three synergies from data on human grasping experiments and trained a neural network with the features of the objects and the coefficients of the synergies. Then, the trained neural network was employed to control robot grasping via an individually optimized mapping between the human hand and the robot hand. As force control was unavailable on our robot hand, we designed a simple strategy for the robot to grasp and hold the objects by exploiting tactile feedback at the fingers. Experimental results demonstrated that the system can generalize the transferred skills to grasp new objects.     

视频编码参数对目标识别性能影响的研究

国内外研究人员对图像目标分类识别和视频编码传输问题都分别进行了大量研究,但是对于视频编码参数对目标识别性能影响的定量关系,还没有公开的文献报导。针对这一问题,该文选择典型的目标识别算法可变部件模型(DPM)和最常用的视频编码方法H.264/AVC作用测试对象,通过设计的编码和检测实验,研究了码率和分辨率参数对视频目标识别性能的影响,并拟合了识别性能随码率和分辨率变化的函数关系。通过选取编码器合适的码率和分辨率工作参数,可以获得信道带宽与视频目标识别性能的折中,为设计不同视频应用的编码优化目标函数提供了依据。     

基于交叉自注意力机制的LiDAR点云三维目标检测

针对基于深度学习的激光雷达(light detection and ranging, LiDAR)点云三维(3D)目标检测对小目标的检测精度较低和噪声干扰问题,提出一种基于交叉自注意力机制的3D点云目标检测方法CSA-RCNN (cross self-attention region convolutional neural network)。利用交叉自注意力(cross self-attention, CSA)同时学习点云的坐标和特征,并设计多尺度融合(multi-scale fusion, MF)模块自适应捕捉各层级多尺度特征。此外,还设计重叠采样策略对感兴趣目标区域选择性地重采样以获得更多前景点,有效降低了噪声采样。在广泛使用的KITTI数据集上进行算法性能测试,结果表明,本文方法对行人等小目标的检测精度有较大提升,平均精度均值相比PointRCNN等4种经典算法均获得提升,显著提高3D点云目标的检测性能。     

Analyzing of flexible gripper by computational intelligence approach

Adaptive grippers should be able to detect and recognize grasping objects. To be able to do it control algorithm need to be established to control gripper tasks. Compliant underactuated mechanisms with passive behavior can be used for modelling of adaptive robotic fingers. Undearactuation is a feature which allows fully adaptability of robotic fingers for different objects. In this study gripper with two fingers was established. Finite element method (FEM) procedure was used to optimize the gripper structural topology. Kinetostatic model of the underactuated finger mechanism was analyzed. This design of the gripper has embedded sensors as part of its structure. The use of embedded sensors in a robot gripper gives the control system the ability to control input displacement of the gripper and to recognize specific shapes of the grasping objects. Since the conventional control strategy is a very challenging task, soft computing based controllers are considered as potential candidates for such an application. The sensors could be used for grasping shape detection. Given that the contact forces of the finger depend on contact position of the finger and object, it is suitable to make a prediction model for the contact forces in function of contact positions of the finger and grasping objects. The prediction of the contact forces was established by using a soft computing (computational intelligence) approach. To perform the contact forces estimation adaptive neuro-fuzzy (ANFIS) methodology was used. FEM simulations were performed in order to acquire experimental data for ANFIS training. The main goal was to apply ANFIS network in order to find correlation between sensors’ stresses and finger contact forces. Afterwards ANFIS results were compared with benchmark models (extreme learning machine (ELM), extreme learning machine with discrete wavelet algorithm (ELM-WAVELET), support vector machines (SVM), support vector machines with discrete wavelet algorithm (SVM-WAVELET), genetic programming (GP) and artificial neural network (ANN)). The reliability of these computational models was analyzed based on simulation results.     

A CMA-ES-Based Adversarial Attack Against Black-Box Object Detectors

Object detection is one of the essential tasks of computer vision. Object detectors based on the deep neural network have been used more and more widely in safe-sensitive applications, like face recognition, video surveillance, autonomous driving, and other tasks. It has been proved that object detectors are vulnerable to adversarial attacks. We propose a novel black-box attack method, which can successfully attack regression-based and region-based object detectors. We introduce methods to reduce search dimensions, reduce the dimension of optimization problems and reduce the number of queries by using the Covariance matrix adaptation Evolution strategy (CMA-ES) as the primary method to generate adversarial examples. Our method only adds adversarial perturbations in the object box to achieve a precise attack. Our proposed attack can hide the specified object with an attack success rate of 86％ and an average number of queries of 5, 124, and hide all objects with a success rate of 74％and an average number of queries of 6, 154. Our work illustrates the effectiveness of the CMA-ES method to generate adversarial examples and proves the vulnerability of the object detectors against the adversarial attacks.     

基于孪生网络的快速视频目标分割

视频目标分割是计算机视觉领域中的一个研究热点,传统基于深度学习的视频目标分割方法在线微调深度网络,导致分割耗时长,难以满足实时的需求.本文提出一种快速的视频目标分割方法.首先,参数共享的孪生编码器子网将参考流和目标流映射到相同的特征空间,使得相同的目标具有相似的特征.然后,全局特征提取子网在特征空间中匹配给定目标相似的特征,定位目标对象.最后,解码器子网将目标特征还原,并通过连接目标流的低阶特征,提供边缘信息,最终输出目标的分割掩码.在公开基准数据集上的实验表明,本文方法的分割速度有大幅度提升,同时具有较好的分割效果.