基于卷积神经网络的行人重识别算法

发布时间：2018-01-16 11:21

本文关键词：基于卷积神经网络的行人重识别算法　出处：《华东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着监控摄像头在各领域的大量应用,传统的人工监控方法无法应对由此产生的海量监控视频。行人重识别是指在多台摄像机监控下进行行人匹配,即给定一个行人目标,在多台不同位置的摄像机不同时刻拍摄的视频中找到该目标。行人重识别(person re-identification)技术是智能视频分析、视频监控、人机交互等诸多领域的核心技术,已经成为计算机视觉领域的研究热点。但是因为光照、视角、姿势、遮挡和分辨率等因素的影响,使得行人重识别技术存在很大的挑战性。行人重识别通常主要包含两个步骤,首先设计有效的描述行人的特征,然后通过度量学习算法进行相似性度量。传统的行人重识别方法依靠人工设计的特征,但由于同一个行人在不同图像中可能有很大差异,而不同的行人又可能看起来很相像,使得这些手工特征很难应用到复杂的现实环境中。深度学习目前已经成功地应用在计算机视觉的很多领域,如手写字符识别、目标检测、图像分类、人脸识别等,在行人重识别领域也有一定的研究。本文采用深度学习方法对行人重识别进行研究,主要研究内容包括:1、基于卷积神经网络中的分类模型的行人重识别研究。不同于常用的相似性度量中的对比损失和三重损失函数,我们用SoftMax损失训练网络。首先对ImageNet数据库上预训练好的AlexNet用行人数据库进行微调(Fine-Tuning),用该网络提取行人特征,并采用目前效果较好的度量学习方法进行识别。其次设计了一个行人分类专用的卷积神经网络,用于提取行人特征,然后用度量学习算法进行行人重识别。2、提出了一种改进的Siamese结构的基于深度卷积神经网络的行人重识别模型。训练时结合了分类和相似性度量,从而增大类间距离,缩小类内距离,提取出行人的有效特征,然后再进一步用度量学习算法进行相似度度量。本文在三个公开的行人重识别数据集上进行实验,采用累积匹配特征(Cumulative Matching Characteristic,CMC)曲线对实验结果进行验证,与其他行人重识别算法进行对比试验,我们的模型优于大部分现有模型。
[Abstract]:With a large number of applications of surveillance cameras in various fields, the traditional manual monitoring methods can not cope with the resulting mass of surveillance video. Pedestrian recognition refers to pedestrian matching under the surveillance of multiple cameras. Given a pedestrian target. The target is found in videos taken at different times by cameras in different locations. Pedestrian re-identification is intelligent video analysis. Video surveillance, human-computer interaction and other fields of core technology, has become a research hotspot in the field of computer vision, but due to lighting, perspective, posture, occlusion and resolution and other factors. Pedestrian recognition technology has a great challenge. Pedestrian recognition usually consists of two steps: first, design an effective description of pedestrian characteristics. Traditional pedestrian recognition methods rely on artificial features, but the same pedestrian may be very different in different images. However, different pedestrians may look very similar, which makes it difficult to apply these manual features to complex real environment. Depth learning has been successfully applied in many fields of computer vision. Such as handwritten character recognition, target detection, image classification, face recognition, there are also some research in the field of pedestrian recognition. The main research contents include: 1, pedestrian recognition based on classification model in convolutional neural network, which is different from the contrast loss and triple loss function in similarity measurement. We use the SoftMax loss training network. Firstly, we fine-tune the pre-trained AlexNet using pedestrian database on the ImageNet database. The network is used to extract pedestrian features, and the current effective metric learning method is used to identify them. Secondly, a special convolutional neural network for pedestrian classification is designed to extract pedestrian features. Then the measurement learning algorithm is used for pedestrian recognition. 2. An improved Siamese structure pedestrian recognition model based on deep convolution neural network is proposed. The training combines classification and similarity measurement to increase inter-class distance and reduce intra-class distance. The effective features of travelers are extracted, and then the similarity is measured using metric learning algorithm. Experiments are carried out on three published pedestrian recognition data sets. The experimental results were verified by cumulative Matching character curve. Compared with other pedestrian recognition algorithms, our model is superior to most existing models.
【学位授予单位】：华东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】