深度学习行人检测方法综述

行人检测技术在智能交通系统、智能安防监控和智能机器人等领域均表现出了极高的应用价值,已经成为计算机视觉领域的重要研究方向之一。得益于深度学习的飞速发展,基于深度卷积神经网络的通用目标检测模型不断拓展应用到行人检测领域,并取得了良好的性能。但是由于行人目标内在的特殊性和复杂性,特别是考虑到复杂场景下的行人遮挡和尺度变化等问题,基于深度学习的行人检测方法也面临着精度及效率的严峻挑战。本文针对上述问题,以基于深度学习的行人检测技术为研究对象,在充分调研文献的基础上,分别从基于锚点框、基于无锚点框以及通用技术改进(例如损失函数改进、非极大值抑制方法等)3个角度,对行人检测算法进行详细划分,并针对性地选取具有代表性的方法进行详细结合和对比分析。本文总结了当前行人检测领域的通用数据集,从数据构成角度分析各数据集应用场景。同时讨论了各类算法在不同数据集上的性能表现,对比分析各算法在不同数据集中的优劣。最后,对行人检测中待解决的问题与未来的研究方法做出预测和展望。如何缓解遮挡导致的特征缺失问题、如何应对单一视角下尺度变化问题、如何提高检测器效率以及如何有效利用多模态信息提高行人检测精度,均是值得进一步...     

Full-scaled deep metric learning for pedestrian re-identification

Multimedia Tools and Applications - The pedestrian re-identification problem (i.e., re-id) is essential and pre-requisite in multi-camera video surveillance studies, provided the fact that...     

基于半监督学习的行人检测方法研究

本文提出了基于半监督学习的行人检测方法,用以解决大量的无标记样本问题。在集成分类器的训练过程中,选择BP神经网络分类器、SVM分类器和KNN分类器作为3个子分类器,利用协同训练机制对各个子分类器进行协同训练。针对半监督学习中误标记样本问题,引入富信息策略和辅助学习策略消除训练过程引入的噪声,同时充分利用无标记样例,进而提高分类器的分类精度。通过对测试集和实时视频进行的行人检测实验,证明了本文方法的可行性和有效性。     

基于迁移学习的行人检测研究进展

行人检测是计算机视觉的研究热点和难点,近年来基于机器学习的行人检测技术取得了长足的进步,但由于不同场景的数据分布存在差异,已有检测器在新场景下的行人检测性能出现显著下降。为了避免繁琐的人工标注,充分利用原有检测器和标注样本,基于迁移学习的行人检测研究受到越来越多的关注。对其中涉及到的样本获取、迁移学习机制等关键技术进行综述,并从多个角度对现有方法进行分析和比较,最后对该技术的未来进行展望。     

Scene-specific pedestrian detection based on transfer learning and saliency detection for video surveillance

Pedestrian detection is a fundamental problem in video surveillance and has achieved great progress in recent years. However the performance of a generic pedestrian detector trained on some public datasets drops significantly when it is applied to some specific scenes due to the difference between source training samples and pedestrian samples in target scenes. We propose a novel transfer learning framework, which automatically transfers a generic detector to a scene-specific pedestrian detector without manually labeling training samples from target scenes. In our method, we get initial detected results and several cues are used to filter target templates whose labels we are sure about from the initial detected results. Gaussian mixture model (GMM) is used to get the motion areas in each video frame and some other target samples. The relevancy between target samples and target templates and the relevancy between source samples and target templates are estimated by sparse coding and later used to calculate the weights for source samples and target samples. Saliency detection is an essential work before the relevancy computing between source samples and target templates for eliminating interference of non-salient region. We demonstrate the effectiveness of our scene-specific detector on a public dataset, and compare with the generic detector. Detection rates improves significantly, and also it is comparable with the detector trained by a lot of manually labeled samples from the target scene.     

基于轻量化深度学习网络的工业环境小目标缺陷检测

工业环境下表面缺陷检测是质量管理的重要一环,具有重要的研究价值.通用检测网络(如YOLOv4)已被证实在多种数据集检测方面是有效的,但是在工业环境的缺陷检测仍需要解决两个问题:一是缺陷实例在表面占比过小,属于典型的小目标检测问题;二是通用检测网络结构复杂,很难部署在移动设备上.针对上述问题,提出一种基于轻量化深度学习网络的工业环境小目标缺陷检测方法.应用GhostNet替代YOLOv4主干特征提取网络,提高网络特征提取能力及降低算法复杂度,并通过改进式PANet结构增加YOLO预测头中高维特征图比例以实现更好的性能.以发动机金属表面缺陷检测为例进行实验分析,结果表明该模型在检测精度(mAP)提升5.83%的同时将网络模型参数量降低83.5%,检测速度提升2倍,同时满足缺陷检测的精度和实时性要求.     

基于深度学习的行人轨迹预测方法综述

为了规划合理的路径以规避行人,针对行人轨迹预测的研究具有广泛的应用价值.基于手工特征的传统方法难以预测复杂场景下的行人轨迹.深度学习以人工神经网络为架构,具有强大的学习能力,在各个领域取得了显著的效果.基于深度学习的行人轨迹预测方法已逐渐发展为一种趋势.为了宏观把握基于深度学习的行人轨迹预测的研究状况,首先,对不同方法进行组织与分类,比较不同方法的优缺点,讨论不同方法在行人轨迹预测领域的应用与发展;其次,根据行人轨迹预测模型的设计差异,对比不同算法对模型性能产生的影响;最后,针对行人轨迹预测中存在的问题,对基于深度学习的行人轨迹预测方法的未来发展进行了展望.     

PSC-Net:learning part spatial co-occurrence for occluded pedestrian detection

Detecting pedestrians,especially under heavy occlusion,is a challenging computer vision prob-lem with numerous real-world applications.This paper introduces a n...     

深度学习目标检测方法综述

目标检测的任务是从图像中精确且高效地识别、定位出大量预定义类别的物体实例。随着深度学习的广泛应用,目标检测的精确度和效率都得到了较大提升,但基于深度学习的目标检测仍面临改进与优化主流目标检测算法的性能、提高小目标物体检测精度、实现多类别物体检测、轻量化检测模型等关键技术的挑战。针对上述挑战,本文在广泛文献调研的基础上,从双阶段、单阶段目标检测算法的改进与结合的角度分析了改进与优化主流目标检测算法的方法,从骨干网络、增加视觉感受野、特征融合、级联卷积神经网络和模型的训练方式的角度分析了提升小目标检测精度的方法,从训练方式和网络结构的角度分析了用于多类别物体检测的方法,从网络结构的角度分析了用于轻量化检测模型的方法。此外,对目标检测的通用数据集进行了详细介绍,从4个方面对该领域代表性算法的性能表现进行了对比分析,对目标检测中待解决的问题与未来研究方向做出预测和展望。目标检测研究是计算机视觉和模式识别中备受青睐的热点,仍然有更多高精度和高效的算法相继提出,未来将朝着更多的研究方向发展。     

基于深度学习的Webshell检测

以AWD攻防中Webshell检测为背景,在超空间利用模糊C均值聚类分析发现了攻击向量全局稀疏、局部紧密的特点,提出了2种深度学习模型。由于GitHub收集的攻击行为多为随机获取,没有很好的针对性,所以对训练数据的长度进行了限制,并保留了有限的相关样本数量。由于一次攻击与相邻的2～4次操作紧密相关,而且攻击向量垂直方向关联特征明显,水平方向相对稳定,考虑到特征向量在传递过程中规模会减小,增加了卷积层的补零选项。针对深度学习训练曲线中的锯齿振荡现象,证明了Adam优化算法的快速计算公式,并修正了学习参数,不断消除了训练的Loss曲线中的锯齿,使得训练曲线按照指数规律平滑下降,迅速得到需要的训练结果。将目前已有的类似工作与提出的2种深度学习模型进行对比。实验结果表明,提出的的深度学习模型能够很好地检测出AWD中的Webshell攻击。     

High-performance and deep pedestrian detection based on estimation of different parts

Pedestrian detection, despite the recent advances, still is of a great challenge to computer vision in wide range of diversified applications such as urban autonomous driving and intelligent transportation. Deep convolutional neural network has greatly contributed to the recent advances in pedestrian detection algorithms. The aim of this paper is to use modified single-shot detector (SSD) approach in pedestrian detection and then improve it by a novel deep architecture. The proposed deep architecture extracts initial Region of Interests (RoIs) using SSD approach, while it employs nine parallel fast RCNNs based on inception modules to estimate nine different parts of body. The proposed method takes the advantage of a secure border in each initial RoI to both create an Extended Region of Candidate Pedestrian (ERCP) and also to extract multi-RoIs. It then selects a number of RoIs within the ERCP as detected pedestrians which satisfy few reasonable criteria. We also propose a new training approach based on different body parts estimation which searches the best RoIs. Comprehensive experimental results demonstrate that the proposed method, deep model based on parts in pedestrian proposals, is a highly effective method that achieves very competitive performance on two most popular pedestrian detection datasets: Caltech-USA and INRIA. We have improved the log-average miss rate on the Caltech-USA and INRIA pedestrian datasets to 7.28% and 4.96%, respectively.

     

集成学习分布式异常检测方法

研究了基于模型共享的集成学习分布式异常检测模型,采用多数投票、边界扩展、平均叠加和距离加权4种不同的集成学习方法得到全部的局部模型;采用交换本地数据挖掘模型的方式来实现数据共享,从而构造出一个总体的集成学习模型。从全局的观点检测异常,减少了集中式检测所需数据的传输量,有效保护了数据提供者的隐私性。仿真实验结果表明,该方法的检测性能与集中式检测的性能相当,甚至更好。     

TensorFlow平台下基于深度学习的数字识别

TensorFlow是谷歌开源的机器学习及深度学习框架,具有高度的灵活性,可以运行在多种平台上,如CPU、GPU以及移动设备,支持当前流行的深度学习模型。卷积神经网络具有多个处理层,能对图像的特征进行逐层抽象,相比于传统的图像识别方法具有良好的效果,对输入图像的旋转、扭曲、变形具有良好的鲁棒性,并且不用对图像进行预处理,简化了图像识别的步骤。在TensorFlow平台上,搭建了一个卷积神经网络模型,利用MNIST数据集对模型进行训练及测试,最终测试能达到99%的识别率。     

Fusion of multispectral data through illumination-aware deep neural networks for pedestrian detection

Multispectral pedestrian detection has received extensive attention in recent years as a promising solution to facilitate robust human target detection for around-the-clock applications (e.g., security surveillance and autonomous driving). In this paper, we demonstrate illumination information encoded in multispectral images can be utilized to boost the performance of pedestrian detection significantly. A novel illumination-aware weighting mechanism is present to depict illumination condition of a scene accurately. Such illumination information is incorporated into two-stream deep convolutional neural networks to learn multispectral human-related features under different illumination conditions (daytime and nighttime). Moreover, we utilized illumination information together with multispectral data to generate more accurate semantic segmentation which is used to supervise the training of pedestrian detector. Putting all of the pieces together, we present an effective framework for multispectral pedestrian detection based on multi-task learning of illumination-aware pedestrian detection and semantic segmentation. Our proposed method is trained end-to-end using a well-designed multi-task loss function and outperforms state-of-the-art approaches on KAIST multispectral pedestrian dataset.     

Intrusion detection by machine learning for multimedia platform

Multimedia Tools and Applications - The multimedia service company, Netflix, increased the number of new subscribers during the Coronavirus pandemic age. Intrusion detection systems for multimedia...     

Application of face image detection based on deep learning in privacy security of intelligent cloud platform

Multimedia Tools and Applications - The application of deep learning-based face detection in the privacy and security of intelligent cloud platforms is studied, in order to resolve the risk of...     

基于深度学习的ARM平台实时人脸识别

针对ARM平台上人脸识别实时性不强和识别率低的问题，提出一种基于深度学习的实时人脸识别方法。首先基于MTCNN人脸检测算法设计了一种实时检测并追踪人脸的算法；然后在ARM平台上基于深度残差网络（ResNet）设计人脸特征提取网络；最后针对ARM平台的特点，使用Mali-GPU加速人脸特征提取网络的运算，分担CPU负荷，提高系统整体运行效率。算法部署在基于ARM的瑞芯微RK3399开发板上，运行速度达到22 帧/s。实验结果表明，与MobileFaceNet相比，该方法在MegaFace上的识别率提升了11个百分点。     

Cross-border e-commerce platform for commodity automatic pricing model based on deep learning

Electronic Commerce Research - With the innovation of information technology and the rise of the Internet economy, cross-border e-commerce has grown up to be an important means and strategy for...     

Malware detection based on deep learning algorithm

In this study we represent malware as opcode sequences and detect it using a deep belief network (DBN). Compared with traditional shallow neural networks, DBNs can use unlabeled data to pretrain a multi-layer generative model, which can better represent the characteristics of data samples. We compare the performance of DBNs with that of three baseline malware detection models, which use support vector machines, decision trees, and the k-nearest neighbor algorithm as classifiers. The experiments demonstrate that the DBN model provides more accurate detection than the baseline models. When additional unlabeled data are used for DBN pretraining, the DBNs perform better than the other detection models. We also use the DBNs as an autoencoder to extract the feature vectors of executables. The experiments indicate that the autoencoder can effectively model the underlying structure of input data and significantly reduce the dimensions of feature vectors.