基于深度特征表示的行人识别方法研究

发布时间：2019-02-08 19:01

【摘要】：行人目标检测与识别技术在智能交通和智能监控系统等领域中一直有着重要应用。虽然该技术在过去几十年中已经取得了长足的发展,但是由于行人外观及复杂的背景环境等因素的影响,行人识别技术仍然是计算机视觉领域公开的难题。针对这一难题,本文提出了新颖的行人识别方法,在3个行人数据集上通过与同类方法比较,证明了所提出方法具有较高的识别率。针对行人识别中的特征表示问题,提出一种混合结构的分层特征表示方法,这种混合结构结合了具有表示能力的词袋结构和学习适应性的深度分层结构。首先利用基于梯度的HOG局部描述符提取局部特征,再通过一个由空间聚集受限玻尔兹曼机(RBM)组成的深度分层编码方法进行编码。对于每个编码层,利用稀疏性和选择性正则化进行无监督受限玻尔兹曼机学习,再应用监督微调来增强分类任务中视觉特征表示,采用最大池化和空间金字塔方法得到高层图像特征表示。最后采用线性支持向量机进行行人识别,提取深度分层特征遮挡等与目标无关部分自然分离,有效提高了后续识别的准确性。针对最大池化方法的局限性,本文在此基础上采用正负最大池化方法进行行人识别。正负最大池化方法克服了最大池化方法中编码系数非负的问题,使得识别准确率进一步提高,但实时性依然未能有效解决。最后总结了全文的研究工作,指出行人识别所面临的问题和难点,并对未来需要完善的相关问题进行了展望。
[Abstract]:Pedestrian target detection and recognition technology has been an important application in intelligent traffic and intelligent monitoring systems. Although the technology has made great progress in the past few decades, due to the influence of pedestrian appearance and complex background environment, pedestrian recognition technology is still an open problem in the field of computer vision. In order to solve this problem, a novel pedestrian recognition method is proposed in this paper. Comparing with the similar methods on three pedestrian datasets, it is proved that the proposed method has a high recognition rate. In order to solve the problem of feature representation in pedestrian recognition, a hierarchical feature representation method for hybrid structures is proposed. This hybrid structure combines the lexical bag structure with the expressive ability and the depth stratified structure with learning adaptability. Firstly, the local feature is extracted by using the gradient-based HOG local descriptor, and then the local feature is encoded by a depth stratified coding method composed of a spatial aggregation constrained Boltzmann machine (RBM). For each coding layer, sparse and selective regularization are used for unsupervised constrained Boltzmann machine learning, and then supervised fine-tuning is applied to enhance visual feature representation in classification tasks. The method of maximum pool and spatial pyramid is used to obtain the feature representation of high-level images. Finally, linear support vector machine (LSVM) is used for pedestrian recognition, extracting depth stratified feature occlusion and other independent parts of the target, which can effectively improve the accuracy of subsequent recognition. Aiming at the limitation of maximum pool method, this paper adopts positive and negative maximum pool method to identify pedestrians. The positive and negative maximum pool method overcomes the problem of non-negative coding coefficient in the maximum pool method, which improves the recognition accuracy further, but the real-time performance is still not solved effectively. Finally, the paper summarizes the research work, points out the problems and difficulties faced by pedestrian recognition, and looks forward to the related problems that need to be improved in the future.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】