煤矿探水卸杆动作识别研究

针对煤矿井下探水作业监工人员通过观看视频来监控卸杆作业的方式存在效率低下且极易出错的问题,提出利用三维卷积神经网络(3DCNN)模型对探水作业中的卸杆动作进行识别。3DCNN模型使用3D卷积层自动完成动作特征提取,通过3D池化层对运动特征进行降维,通过Softmax分类处理来识别卸杆动作,并使用批量归一化层提高模型的收敛速度和识别准确率。采用3DCNN模型对卸杆动作进行识别时,首先对数据集进行预处理,从每段视频中均匀抽取几帧图像作为某动作的代表,并降低分辨率;然后采用训练集对3DCNN模型进行训练,并保存训练好的权重文件;最后采用训练好的3DCNN模型对测试集进行测试,得出分类结果。实验结果表明,设置采样帧数为10帧、分辨率为32×32、学习率为0.000 1,3DCNN模型对卸杆动作的识别准确率最高可达98.86%。     

基于深度学习的人体动作识别方法

对基于机器视觉的人体动作识别的成果进行研究,为提高视频数据集中人体动作的识别率,提出一种改进的深度网络模型。采用稠密光流方法处理数据,结合二维卷积神经网络(2DCNN)、三维卷积神经网络(3DCNN)和长短期记忆神经网络(LSTM)对动作特征进行提取,利用Softmax分类器识别分类。通过KTH数据集进行实验对比验证,分析结果表明,改进模型相比其它已有模型具有更高的识别率,动作识别效果更优。     

一种改进的GRU-InFCN人体行为识别模型

为了克服传统机器学习方法在采用传感器数据进行人体行为识别领域上识别效果对人工特征选取依赖严重、识别准确率不高等问题,提出一种改进的全卷积神经网络和多层循环神经网络并联的深度学习模型(GRU-InFCN),并对传感器数据特征进行自动提取,实现人体动作的识别。该模型通过多尺度卷积神经网络和双层GRU网络(Gated Recurrent Unit,GRU)分别对传感器数据进行特征提取,将特征矩阵在矩阵维度上进行特征拼接再通过Softmax完成特征分类。实验结果表明,在开源人体行为识别(HAR)数据集上采用该方法进行人体行为识别,准确率达到了97.76%。该模型在取得高准确率的同时,避免了复杂的信号预处理和特征工程。     

基于深度卷积神经网络的面部表情识别研究

本研究为了解决传统面部表情识别模型准确率较低的问题，基于深度卷积神经网络（Deep Convolutional Neural Network,DCNN）提出一种新的改进神经网络模型，与传统模型相对比，本模型将其核心的卷积层替换成了深度可分离卷积层，同时搭配卷积残差块的使用，使网络能够有效减少参数的情况下，能够提取多尺度上的特征信息，从而有效地保留了细节特征。最后通过仿真对比，证明本研究提出的DCNN网络具有突出的性能特点，适合用于面部表情识别任务。     

基于视频深度学习的时空双流人物动作识别模型

深度学习在人物动作识别方面已取得较好的成效,但当前仍然需要充分利用视频中人物的外形信息和运动信息。为利用视频中的空间信息和时间信息来识别人物行为动作,提出一种时空双流视频人物动作识别模型。该模型首先利用两个卷积神经网络分别抽取视频动作片段空间和时间特征,接着融合这两个卷积神经网络并提取中层时空特征,最后将提取的中层特征输入到3D卷积神经网络来完成视频中人物动作的识别。在数据集UCF101和HMDB51上,进行视频人物动作识别实验。实验结果表明,所提出的基于时空双流的3D卷积神经网络模型能够有效地识别视频人物动作。     

多模态零样本人体动作识别

目的 在人体行为识别算法的研究领域,通过视频特征实现零样本识别的研究越来越多。但是,目前大部分研究是基于单模态数据展开的,关于多模态融合的研究还较少。为了研究多种模态数据对零样本人体动作识别的影响,本文提出了一种基于多模态融合的零样本人体动作识别（zero-shot human action recognition framework based on multimodel fusion, ZSAR-MF）框架。方法 本文框架主要由传感器特征提取模块、分类模块和视频特征提取模块组成。具体来说,传感器特征提取模块使用卷积神经网络（convolutional neural network, CNN）提取心率和加速度特征;分类模块利用所有概念（传感器特征、动作和对象名称）的词向量生成动作类别分类器;视频特征提取模块将每个动作的属性、对象分数和传感器特征映射到属性—特征空间中,最后使用分类模块生成的分类器对每个动作的属性和传感器特征进行评估。结果 本文实验在Stanford-ECM数据集上展开,对比结果表明本文ZSAR-MF模型比基于单模态数据的零样本识别模型在识别准确率上提高了4 %左右。结论 本文所提出的基于多模态融合的零样本人体动作识别框架,有效地融合了传感器特征和视频特征,并显著提高了零样本人体动作识别的准确率。     

基于数据融合的CNN方法用于人体活动识别

针对人体活动识别,现有的研究更多关注于传感器之间的融合,较少将研究方向转移到三轴加速度计等传感器内轴之间的相关性方面。为有效利用轴之间的相关性,提出一种基于数据融合的卷积神经网络方法。借助利用轴之间的相关性的单通道数据融合方法得到融合数据,将融合数据输入到卷积神经网络中提取特征。在WISDM数据集上的实验结果表明,该方法的准确率达到了98.80%,优于不使用数据融合的卷积神经网络方法。     

基于三轴加速度传感器的手势识别

针对手势交互中手势信号的相似性和不稳定性,设计实现一种基于三轴加速度传感器的手势识别方案。采用MMA7260加速度传感器采集主手腕的手势动作信号,根据手势加速度信号的特点,进行手势动作数据窗口的自动检测、信号去噪和重采样等预处理,通过提取手势动作的关键特征,构造离散隐马尔可夫模型,实现手势动作识别。实验结果证明该方案的识别精度较高。     

基于9轴姿态传感器的CNN旗语动作识别方法

区别于传统光纤传感器、图像识别和Kinect深度图像的旗语动作识别方法,提出了一种基于9轴姿态传感器的旗语动作识别方法.该方法通过佩戴在手腕处的9轴姿态传感器来采集旗语动作的3轴加速度、3轴角速度以及3轴磁偏角数据;在运用卷积神经网络(Convolutional Neural Network,CNN)分类模型的基础上对其中的数据进行预处理,并通过分类识别算法对其进行改进;在数据预处理阶段,利用小波分解与重构函数对采集到的9轴数据进行高频去噪和低频信息提取,通过时间序列加窗进行分割处理,对各动作样本进行维度和长度统一;在特征提取阶段,采用构建的双卷积层、单池化层、单全连接层网络模型对重构数据进行特征提取;在分类识别阶段,提出一种CrossEntropy-Logistic联合损失函数来对5种动作进行迭代训练.实验结果表明,所提方法利用detcoef小波分解与重构函数对信号进行低频细节系数提取并采用一维CNN对降噪后的数据进行特征提取,通过CL联合损失函数对预测损失值和预测概率进行融合,分析所得到的训练准确率与测试准确率,在与各类方法的对比中取得了最高值,其平均训练识别率可达99％以上,测试准确率可达94％.     

基于双通道卷积神经网络的航班延误预测模型

针对航班延误预测数据量大、特征提取困难而传统算法处理能力有限的问题,提出一种基于双通道卷积神经网络（DCNN）的航班延误预测模型。首先,该模型将航班数据和气象数据进行融合,应用DCNN进行自动特征提取,采用批归一化（BN）和Padding策略优化,提升到港延误等级的分类预测性能;然后,在卷积神经网络（CNN）基础上加入直通通道,以保证特征矩阵的无损传输,增强深度网络的畅通性;同时引入卷积衰减因子对卷积通道的特征矩阵进行稀疏性限制,控制不同网络深度的特征叠加比例,维持模型的稳定性。实验结果表明,所提模型与传统模型相比,具有更强的数据处理能力。通过数据融合,航班延误预测准确率可提高1个百分点;加深网络深度后,该模型能保证梯度的稳定,从而训练更深的网络,使准确率提升至92.1%。该基于DCNN算法的模型特征提取充分,预测性能优于对比模型,可更好地服务于民航决策。     

带有协方差矩阵的卷积神经网络在人体运动识别中的应用

目前,深度学习已经在各种人体运动识别（HAR）任务中发挥了重要作用。但是,由于运动数据具有时间序列和包含肢体动作的特殊性,现有神经网络在进行卷积操作时会导致数据高度相关,并且随着网络影响到下一层,这限制了模型的识别效果。为此,提出了一种带有协方差矩阵的改进卷积神经网络用于HAR场景,通过矩阵变换搭建一种去相关的网络结构来消除相关性问题,可以在网络表现不佳时替代现有的批量归一化（BN）层用于归一化数据。在4个HAR公共数据集上进行实验,并与传统CNN和带有BN层的模型进行比较。实验结果表明,对比此前的深度学习网络,改进的神经网络有1%~2%的性能提升,验证了该方法的有效性,并将程序移植到了移动端进行实时运动识别。     

Online continual learning for human activity recognition

Sensor-based human activity recognition (HAR), with the ability to recognise human activities from wearable or embedded sensors, has been playing an important role in many applications including personal health monitoring, smart home, and manufacturing. The real-world, long-term deployment of these HAR systems drives a critical research question: how to evolve the HAR model automatically over time to accommodate changes in an environment or activity patterns. This paper presents an online continual learning (OCL) scenario for HAR, where sensor data arrives in a streaming manner which contains unlabelled samples from already learnt activities or new activities. We propose a technique, OCL-HAR, making a real-time prediction on the streaming sensor data while at the same time discovering and learning new activities. We have empirically evaluated OCL-HAR on four third-party, publicly available HAR datasets. Our results have shown that this OCL scenario is challenging to state-of-the-art continual learning techniques that have significantly underperformed. Our technique OCL-HAR has consistently outperformed them in all experiment setups, leading up to 0.17 and 0.23 improvements in micro and macro F1 scores.     

A time-efficient convolutional neural network model in human activity recognition

Activity recognition represents the task of classifying data derived from different sensor types into one of predefined activity classes. The most popular and beneficial sensors in the area of action recognition are inertial sensors such as accelerometer and gyroscope. Convolutional neural network (CNN) as one of the best deep learning methods has recently attracted much attention to the problem of activity recognition, where 1D kernels capture local dependency over time in a series of observations measured at inertial sensors (3-axis accelerometers and gyroscopes) while in 2D kernels apart from time dependency, dependency between signals from different axes of same sensor and also over different sensors will be considered. Most convolutional neural networks used for recognition task are built using convolution and pooling layers followed by a few number of fully connected layers but large and deep neural networks have high computational costs. In this paper, we propose a new architecture that consists solely of convolutional layers and find that with removing the pooling layers and instead adding strides to convolution layers, the computational time will decrease notably while the model performance will not change or in some cases will even improve. Also both 1D and 2D convolutional neural networks with and without pooling layer will be investigated and their performance will be compared with each other and also with some other hand-crafted feature based methods. The third point that will be discussed in this paper is the impact of applying fast fourier transform (FFT) to inputs before training learning algorithm. It will be shown that this preprocessing will enhance the model performance. Experiments on benchmark datasets demonstrate the high performance of proposed 2D CNN model with no pooling layers.

     

Design of data feature-driven 1D/2D convolutional neural networks classifier for recycling black plastic wastes through laser spectroscopy

In this study, two types of convolutional neural network (CNN) classifiers are designed to handle the problem of classifying black plastic wastes. In particular, the black plastic wastes have the property of absorbing laser light coming from spectrometer. Therefore, the classification of black plastic wastes remains still a challenging problem compared to classifying other colored plastic wastes using existing spectroscopy (i.e., NIR). When it comes the classification problem of black plastic wastes, effective classification techniques by the laser spectroscopy of Fourier Transform-Infrared Radiation (FT-IR) with Attenuated Total Reflectance (ATR) and Raman to analyze the classification problem of black plastic wastes are introduced. Due to the strong ability of extracting spatial features and remarkable performance in image classification, 1D and 2D CNN through data features are designed as classifiers. The technique of chemical peak points selection is considered to reduce data redundancy. Furthermore, through the selection of data features based on the extracted 1D data with peak points is introduced. Experimental results demonstrate that 2DCNN classifier designed with the help of 2D data feature selection as well as 1DCNN classifier shows the best performance compared with other reported methods for classifying black plastic wastes.     

Performance-enhanced real-time lifestyle tracking model based on human activity recognition (PERT-HAR) model through smartphones

The identification of human actions and their representation and categorization in an automated system through training and learning is considered the human activity recognition (HAR) process. Tracking systems capture and read human actions ranging from the identification of plain movements to the comprehension of habits and practices. As such, HARs have found their use in areas such as health care, with a special focus on elderly patients’ care, safety arrangements and supervision areas and in applications designed for smart homes. Sensor and visual devices enable HAR, and there is a multitude of sensor classifications, such as sensors that can be worn, sensors tagged to a target and sensors tagged to the background. The automated learning methodologies in HAR are either handcrafted or deep learning or a combination of both. Handcrafted models can be regional or wholesome recognition models such as RGB, 3D mapping and skeleton data models, and deep learning models are categorized into generative models such as LSTM (long short-term memory), discriminative models such as convolutional neural networks (CNNs) or a synthesis of such models. Several datasets are available for undertaking HAR analysis and representation. The hierarchy of processes in HAR is classified into gathering information, preliminary processing, property derivation and guiding based on framed models. The proposed study considers the role of smartphones in HARs with a particular interest in keeping a tab on the lifestyle of subjects. Smartphones act as HAR devices with inbuilt sensors with custom-made applications, and the merits of both handcrafted and deep learning models are considered in framing a model that can enable lifestyle tracking in real time. This performance-enhanced real-time tracking human activity recognition (PERT-HAR) model is economical and effective in accurate identification and representation of actions of the subjects and thereby provides more accurate data for real-time investigation and remedial measures. This model achieves an accuracy of 97–99% in a properly controlled environment.

     

Saliency-based deep convolutional neural network for no-reference image quality assessment

In this paper, we proposed a novel method for No-Reference Image Quality Assessment (NR-IQA) by combining deep Convolutional Neural Network (CNN) with saliency map. We first investigate the effect of depth of CNNs for NR-IQA by comparing our proposed ten-layer Deep CNN (DCNN) for NR-IQA with the state-of-the-art CNN architecture proposed by Kang et al. (2014). Our results show that the DCNN architecture can deliver a higher accuracy on the LIVE dataset. To mimic human vision, we introduce saliency maps combining with CNN to propose a Saliency-based DCNN (SDCNN) framework for NR-IQA. We compute a saliency map for each image and both the map and the image are split into small patches. Each image patch is assigned with a patch importance value based on its saliency patch. A set of Salient Image Patches (SIPs) are selected according to their saliency and we only apply the model on those SIPs to predict the quality score for the whole image. Our experimental results show that the SDCNN framework is superior to other state-of-the-art approaches on the widely used LIVE dataset. The TID2008 and the CISQ image quality datasets are utilised to report cross-dataset results. The results indicate that our proposed SDCNN can generalise well on other datasets.     

基于无线体域网的在线人体活动识别

基于智能手机传感器的人体活动识别是普适计算领域的研究热点.为扩展可识别的活动种类,并提高准确率和实时性,提出了由智能手环和智能手机组建无线体域网通过深度神经网络在线识别人体活动的方法.首先,设计由智能手环和智能手机组成的无线体域网的总体框架;然后,对预处理后的传感信号,构造带有Inception结构的卷积神经网络和长短时记忆递归神经网络来分别提取时空域特征,并结合两类网络结构来融合多模态传感数据,离线进行神经网络模型训练;最后,对训练好的神经网络模型进行优化,并部署到智能手机上,在线实时识别人体活动.实验结果表明,本文方法无需手工设计特征,可自动融合各类异构传感数据,更加准确、高效地识别了更多种类的活动.     

CODON: On Orchestrating Cross-Domain Attentions for Depth Super-Resolution

The ready accessibility of high-resolution image sensors has stimulated interest in increasing depth resolution by leveraging paired color information as guidance. Nevertheless, how to effectively exploit the depth and color features to achieve a desired depth super-resolution effect remains challenging. In this paper, we propose a novel depth super-resolution method called CODON, which orchestrates cross-domain attentive features to address this problem. Specifically, we devise two essential modules: the recursive multi-scale convolutional module (RMC) and the cross-domain attention conciliation module (CAC). RMC discovers detailed color and depth features by sequentially stacking weight-shared multi-scale convolutional layers, in order to deepen and widen the network at low-complexity. CAC calculates conciliated attention from both domains and uses it as shared guidance to enhance the edges in depth feature while suppressing textures in color feature. Then, the jointly conciliated attentive features are combined and fed into a RMC prediction branch to reconstruct the high-resolution depth image. Extensive experiments on several popular benchmark datasets including Middlebury, New Tsukuba, Sintel, and NYU-V2, demonstrate the superiority of our proposed CODON over representative state-of-the-art methods.

     

A survey on the use of machine learning methods in context-aware middlewares for human activity recognition

Human activity recognition (HAR) essentially uses (past) sensor data or complex context information for inferring the activities a user performs in his daily tasks. HAR has been extensively studied using different paradigms, such as different reasoning mechanisms, including probabilistic, rule-based, statistical, logical reasoning, or the machine learning (ML) paradigm, to construct inference models to recognize or predict user activities. ML for HAR allows that activities can be recognized and even anticipated through the analysis of collected data from different sensors, with greater accuracy than the other paradigms. On the other hand, context-aware middlewares (CAMs) can efficiently integrate a large number of different devices and sensors. Moreover, they provide a programmable and auto-configurable infrastructure for streamline the design and construction of software solutions in scenarios where lots of sensors and data are their bases, such as ambient intelligence, smart cities, and e-health domains. In this way, the full integration of ML capabilities as services in CAMs can advance the development of software solutions in these domains when ML is necessary, specially for HAR, which is the basis for many scenarios in these domains. In this work, we present a survey for identifying the state-of-the-art in using ML for HAR in CAMs through a systematic literature review (SLR). In our SLR, we worked to answer four research questions: (i) what are the different types of context reasoners available in CAMs; (ii) what are the ML algorithms and methods used for generating models for context reasoning; (iii) which CAMs support data processing in real time; and (iv) what are the HAR scenarios usually tackled by the research works. In our analysis, we observed that, although ML offers viable approaches to construct inference models for HAR using different ML approaches, including batch learning, adaptive learning and data stream learning, there are yet some gaps and research challenges to be tackled, specially on the use of data stream learning considering concept drift on data, mechanisms for adapting the inference models, and further considering all of this as services in CAMs, specially for HAR.

     

Human Activity Recognition Based on Parallel Approximation Kernel K-Means Algorithm

Recently, owing to the capability of mobile and wearable devices to sense daily human activity, human activity recognition (HAR) datasets have become a large-scale data resource. Due to the heterogeneity and nonlinearly separable nature of the data recorded by these sensors, the datasets generated require special techniques to accurately predict human activity and mitigate the considerable heterogeneity. Consequently, classic clustering algorithms do not work well with these data. Hence, kernelization, which converts the data into a new feature vector representation, is performed on nonlinearly separable data. This study aims to present a robust method to perform HAR data clustering to mitigate heterogeneity in data with minimal resource consumption. Therefore, we propose a parallel approximated clustering approach to handle the computational cost of big data by addressing noise, heterogeneity, and nonlinearity in data using data reduction, filtering, and approximated clustering methods on parallel computing environments that have not been previously addressed. Our key contribution is to treat HAR as big data implemented by approximation kernel K-means approaches and fill the gap between the HAR clustering cost and parallel computing fields. We implemented our approach on Google cloud on a parallel spark cluster, which helped us to process large-scale HAR data across multiple machines of clusters. The normalized mutual information is used as validation metric to assess the quality of the clustering algorithm. Additionally, the precision, recall, f-score metrics values are obtained somehow to compare the results with a classification technique. The experimental results of our clustering approach prove its effectiveness compared with a classification technique and can efficiently detect physical activity and mitigate the heterogeneity of the datasets.