完美论文网提供代写毕业论文、代写本科论文、代写硕士论文、论文发表及查重服务!QQ:505676555

电话:18796337551
当前位置:完美论文网信息系统 → 文章正文

基于YOLOv2多尺度特征融合的步态识别研究

作者:完美论文网  来源:www.wmlunwen.com  发布时间:2019/10/10 9:04:53  

摘要:步态识别是根据人行走的频率、相位以及胳膊摆动等来识别人身份的一种生物识别技术,其特征具有易采集、难以模仿以及可以远距离识别等优点,使得它在安全敏感环境中具有广阔的应用前景。目前在步态识别的相关研究中,基本上都是使用行人检测和跟踪的方法获取步态轮廓图像序列,然后对获取的步态轮廓图像序列进行预处理和周期检测,最后从步态周期图像序列中提取步态特征,并设计相应的分类器进行分类识别。然而在行人检测的过程中容易受背景、行人密集度等因素的影响,以及在分类识别的过程中容易受步态轮廓图像质量、步态特征质量以及视角等因素的影响,因此本文的研究内容主要分为以下三部分:

(1)研究检测行人技术。使用传统的YOLOv2网络训练出来的行人检测模型在背景简单以及行人遮掩不严重的情况下,该模型的检测效果良好,但是当背景复杂以及行人遮掩严重的时候该模型的检测效果较差。针对此问题,本文提出在传统的YOLOv2网络中添加HOG-CSLBP特征提取层,并根据维度聚类方法对INRIA数据集目标聚类分析的结果调整YOLOv2网络的先验框个数与维度值,然后使用改进后网络训练出来的行人检测模型进行行人检测。实验结果表明,在误检率为0.1时该算法的漏检率为9.13%,与传统的YOLOv2网络相比漏检率降低了5.27%。

(2)研究步态图像预处理和周期检测技术。对本文提出的行人检测方法所检测出的行人目标区域,采用背景消减法提取人体轮廓图像并二值化。由于获取的二值轮廓图像中存在着噪声点和空洞,首先采用形态学方法对其进行去噪或者降噪,从而降低对步态识别率的影响;其次本文通过图像归一化方法对经过形态学处理后的步态轮廓图像进行归一化处理,然后对归一化后的步态图像进行周期检测,从而获得行人步态周期图像序列。

(3)研究步态识别技术。将协同表示方法应用于步态识别中可以解决稀疏表示方法计算耗时的问题,但提取步态特征采用的GEI算法没有考虑步态内部轮廓边界信息,导致识别率不高。针对此问题,本文提出使用融合HOG和GEI算法的方法提取步态特征,在此基础上使用协同表示的方法训练,再通过计算测试样本的最小重构误差进行分类。实验结果表明,该方法在单一视角下步态识别准确率平均提高了1.315%,以及跨视角下步态识别准确率平均提高了6.51%。

Gait recognition is a biometricidentification technology that can recognize humans’ identities according tothe frequency, phase and arm swing of humans. With the advantages of easyacquisition, difficult in imitation and remote recognition, these make it hasbroad application prospects in security sensitive environment. In the relatedresearch of gait recognition, the gait contour image sequence is basicallyobtained by the pedestrian detection and tracking method. Then, the acquiredgait contour image sequence is preprocessed and periodically detected. Finally,extract gait features from the gait cycle image sequence and designcorresponding classifier for identity recognition. However, in the process ofpedestrian detection, it is easy to be affected by factors such as backgroundand pedestrian density. And in the process of classification and recognition, itis easy to be affected by factors such as gait contour image quality, gaitfeature quality and viewing angle. Based on the above reasons, this paperlaunches targeted research from three parts:

(1)Research onpedestrian detection technology. In the case of simple background and noserious pedestrian occlusion, the pedestrian detection model based on thetraditional YOLOv2 network training has better detection effect, but thedetection effect of the model is poor when the background is complicated and thepedestrians are seriously concealed. To solve this problem, this paper proposesto add the HOG-CSLBP feature extraction layer to the traditional YOLOv2network, and adjusts the anchor boxes number and dimension value of the YOLOv2network according to the result of clustering analysis of INRIA data set.Finally, the improved network is used to train the pedestrian detection modeland perform pedestrian detection. The experimental result shows that themissing detection rate of the algorithm is 9.13% when the false detection rateis 0.1. Compared with the traditional YOLOv2 network, the missing detectionrate is reduced by 5.27%.

(2)Research on gaitimage preprocessing and period detection technology. For the pedestrian targetarea detected by the pedestrian detection method proposed in this paper,background subtraction method is used to extract human contour image andbinarize it. Since there are noise points and voids in the acquired binarycontour image, the morphological method is first used to denoise or reduce thenoise, thereby reducing the influence of these factors on the gait recognitionrate. Secondly, the gait contour image after morphological processing isnormalized by image normalization method, and then the normalized gait image isperiodically detected to obtain the pedestrian gait periodic image sequence.

(3)Research gaitrecognition technology. Applying collaborative representation method to gaitrecognition research can solve the problem of time-consuming calculation ofsparse representation. While the GEI algorithm for extracting gait featuresdoes not consider gait internal outline information, resulting a lowrecognition rate. To address this problem, this article propose a novel methodto extract gait features by using a fusion HOG and GEI algorithm. On the basisof this method, train the samples by using collaborative representation methodand classify the samples according to the minimum reconstruction error. Theexperimental results show that the accuracy of the proposed method is 1.315%higher than that of the collaborative representation in a single view, and theaccuracy of gait recognition is increased by 6.51% over the cross view.

关键词:步态识别;YOLOv2;HOG-CSLBP;HOG;GEI;协同表示

gaitrecognition;YOLOv2;HOG-CSLBP;HOG;GEI;collaborative representation

联系方式

客服QQ 505676555
客服热线18796337551
网站地址 www.wmlunwen.com
郑重承诺 原创,包修改,包通过!
完美论文网真诚欢迎新老客户的光临与惠顾!