基于人体感知及上下文学习的行人识别方法研究

项目名称： 基于人体感知及上下文学习的行人识别方法研究

项目编号： No.61463017

项目类型： 地区科学基金项目

立项/批准年度： 2015

项目学科： 其他

项目作者： 吕敬钦

作者单位： 华东交通大学

项目金额： 45万元

中文摘要： 行人识别是近几年发展的研究热点，是视频监控、人体或服装检索等的关键技术。当前方法在描述行人方面未深入研究，因而当中的难点尚未解决，如视角和姿态多变、衣着多样、复杂背景干扰等，以及相似度问题；也限制了已有的识别方法难以提升识别率。针对这些难点和问题，并从行人识别的本质出发，本课题基于人体感知和上下文学习来深入研究行人识别方法。主要内容有：1）研究运用语义上下文感知人体模型的方法，以及模型中Semantic Cell的配对方法，以解决前述的主要难点。2）研究人眼视觉识别的机制，以及研究Semantic Cell联合识别的方法，并融入人体配对图模型的优化识别框架中。有望缓解相似度问题，并进一步提升识别率。3）研究基于视频人体模型的多行人跟踪方法，及多视角行人识别方法，将获得更好的识别表现，且适用于繁忙的监控场景。本课题在转变当前行人识别的局限及推动研究深入发展方面，具有重要的意义。

中文关键词： 行人识别；上下文学习；人体感知；多行人跟踪

英文摘要： Pedestrian recognition is a new hot research field in recent years. It's the key technology for video surveillance, body or clothing retrieval and etc. For current methods, the research on body description is not deeply enough. Therefore, the difficulties in this field have not been solved yet, including viewpoint and pose variations, various clothing, background clutters, and the similarity problem. Also improving recognition performance is limited for current methods. Aiming at most difficulties and starting from the nature of pedestrian recognition, we propose to do pedestrian recognition research based on body perceiving and contextual learning. This research mainly includes: 1) semantic context based perceiving method for body model, and matching method for semantic cell, which are expected to solve the difficulties. 2) Study the recognition mechanism of human vision and the idea of recognition by combining multiple semantic cells. And merge both into the optimizing recognition framework of Graphical Model of Body matching, which will alleviate the similarity problem and further raise the recognition rate. 3) Multiple pedestrian tracking method based on video body model; and multiple-view pedestrian recognition method, which are expected to achieve better recognition performance and can be applied to video surveillance in busy scenes. This research will play an important role in changing limitations of current pedestrian recognition and enhancing study progress.

英文关键词： Pedestrian Recognition;Contextual Learning;Body Perceiving;Multiple People Tracking