基于深度学习的行人再识别技术研究

发布时间：2018-03-21 02:44

本文选题：深度神经网络　切入点：行人再识别　出处：《西南交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：行人再识别(Person re-identification)技术是判断在不同监控摄像头下出现的行人图像是否属于同一行人的技术。面对海量增长的监控视频,利用计算机对监控视频中的行人进行再识别的需求应运而生。然而现存的行人再识别算法主要是在已裁剪好的行人图片中匹配查找集和候选集,这是不切实际的,行人的框架在现实考虑中不可能直接给定,目标行人需要在整张图片中被锁定。目前,深度学习在图像识别、语音识别、自然语言处理等多个领域取得了优异的效果。相比于传统人工提取特征的方法,深度神经网络通过从数据中自动学习到更能表征图像的特征并进行分类,更具实际意义。将深度学习应用到行人再识别上已经成为当前的研究热点,但是由于目前行人再识别中如图像分辨率低、遮挡、光照变化等问题使其离实际应用还有很长的距离。本文总结了目前一些行人检测及再识别的常用特征、算法以及深度神经网络结构,并进行深入研究和分析。设计了一种针对端到端行人再识别的预训练网络模型,该模型结合了验证和分类两种网络结构,并利用空间池化操作对不同尺度的输入图片进行特征归一化。在此基础上用性能良好的ResNet-50网络结构对端到端的行人再识别网络结构进行改进。之后在caffe深度学习框架上训练改进的模型并进行多组实验,包括预训练模型的有效性、不同特征维度对网络模型效果的影响、在不同大小的候选集、低分辨率和遮挡子集下的性能分析,以及与当前比较先进的算法进行对比。实验结果证明了本文方法训练出来的模型能够学习到具有较高鲁棒性的特征,大幅度提高了行人再识别的识别率。
[Abstract]:Pedestrian re-identification is a technique to determine whether a pedestrian image under different surveillance cameras belongs to the same pedestrian. The need for rerecognition of pedestrians in surveillance videos arises with the help of computers. However, it is unrealistic for existing pedestrian rerecognition algorithms to match and find sets and candidate sets in a cut pedestrian image. The pedestrian frame cannot be given directly in practical considerations, and the target pedestrian needs to be locked in the entire picture. At present, the depth of learning is in image recognition, speech recognition, Natural language processing and other fields have achieved excellent results. Compared with the traditional methods of artificial feature extraction, depth neural networks can automatically learn from the data to represent the features of images and classify them. The application of depth learning to pedestrian rerecognition has become a hot research topic. However, due to the low image resolution and occlusion in pedestrian rerecognition, This paper summarizes some common features, algorithms and depth neural network structure of pedestrian detection and re-recognition. A pre-training network model for end-to-end pedestrian rerecognition is designed, which combines verification and classification of network structure. The spatial pool operation is used to normalize the input images of different scales. On the basis of this, the end-to-end pedestrian recognition network structure is improved with the good performance of ResNet-50 network structure, and then the caffe depth learning is carried out. Training the improved model on the framework and conducting multiple groups of experiments, Including the effectiveness of the pre-training model, the influence of different feature dimensions on the effectiveness of the network model, and the performance analysis under different sizes of candidate sets, low resolution and occlusion subsets. The experimental results show that the model trained by this method can learn the characteristics of high robustness and greatly improve the recognition rate of pedestrian recognition.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】