变尺度对象追踪技术及应用研究

发布时间：2018-10-29 14:53

【摘要】：随着数字多媒体技术的发展,用户视频、监控视频等多种类型的视频数据总量呈爆炸式增长,为网络内容审查、安防监控等工作带来了极大的挑战。因此,智能视频内容分析系统以及相关视觉技术的研究亟需突破性进展,而其中对象追踪模块作为视频应用的关键步骤,越来越被人们所关注与重视。视觉追踪技术的难点在于:1)需能适应画面光影变化、局部遮挡、背景复杂、相机移动、目标对象尺寸变化等视频噪声,保证准确追踪;2)必须在图像帧之间进行实时在线追踪,不影响视频的播放速度,否则将导致信息传达不畅,或致使追踪精度受损。一个合格的追踪算法,既要具备足够的稳定性与鲁棒性,又需要能够实时高速运行,防止追踪延迟或失效。为了应对上述难点,本论文设计并实现了单滤波变尺度追踪算法。该方法以核化相关单滤波追踪器为基础,利用判决模型对候选窗口进行密集采样所得到的循环结构,提出使用“四角型”高斯类标矩阵作为追踪滤波器的响应训练标注,并从边缘追踪和分块追踪两个角度说明其有效性。不同于已有的多尺度追踪方法,本方法无需在单帧中多次运行滤波操作或多次采样候选窗口,仅通过一次滤波运算即可同时推断出目标对象的新位置与新尺寸,在超实时追踪的前提下,解决单滤波追踪算法无法处理的对象尺寸变化问题。在追踪器执行滤波操作得到“四峰值”响应矩阵后,本论文采用贝叶斯框架推断目标对象所在窗口的具体坐标与边界框长宽。该方法首先从时间与空间两个角度衡量变尺度单滤波器各个局部的追踪置信度并计算结果权值,而后对各分块的追踪结果进行有偏融合,以置信度较高的局部响应来推断目标对象的最终状态,同时抑制有可能出错的追踪结果所造成的影响,相比于原始KCF追踪器的单峰值定位策略而言更具鲁棒性。最后本论文设计并开发了一个小型的监控视频分析与预警系统,该系统分为三个部分:1)基于Faster R-CNN的对象检测模块,在输入的视频文件或流数据中识别兴趣对象;2)变尺度单滤波对象追踪算法,获取目标在每一帧的位置与尺寸;3)具体的异常行为识别模块,根据实际应用要求对目标对象的行为进行分析,如果发现所指定的异常情况则做出警示。
[Abstract]:With the development of digital multimedia technology, the amount of video data such as user video, surveillance video and other kinds of video data increases explosively, which brings great challenges to the work of network content censorship, security monitoring and so on. Therefore, the research of intelligent video content analysis system and related visual technology is in urgent need of breakthrough, and the object tracking module, as a key step of video application, has been paid more and more attention to. The difficulties of visual tracking technology are as follows: 1) it needs to be able to adapt to the video noise such as the change of picture light and shadow, local occlusion, complex background, camera moving and object size change, so as to ensure accurate tracking; 2) real-time online tracking must be carried out between image frames, which will not affect the speed of video playback, otherwise, it will lead to poor communication of information or damage of tracking accuracy. A qualified tracking algorithm not only has sufficient stability and robustness, but also needs to be able to run at high speed in real time to prevent tracking delay or failure. In order to deal with the above difficulties, a single filter variable scale tracking algorithm is designed and implemented in this paper. Based on the kernel correlation single filter tracker and the circular structure obtained by dense sampling of candidate window by decision model, this method proposes to use the "tetragonal" Gao Si class matrix as the response training label of the tracking filter. The effectiveness of the method is illustrated from two aspects: edge tracking and block tracking. Different from the existing multi-scale tracking method, the new position and size of the target object can be inferred by only one filter operation, which does not need to run multiple filtering operations or multiple sampling candidate windows in a single frame. On the premise of super real time tracking, the problem of object size change which can not be handled by single filter tracking algorithm is solved. After the tracker performs the filtering operation to get the "four-peak" response matrix, the Bayesian framework is used to infer the specific coordinates of the target window and the length and width of the boundary frame. This method firstly measures the local tracking confidence of variable scale single filter from time and space angles and calculates the weight of the result. Then the tracking results of each block are partially fused. The local response with high confidence can be used to infer the final state of the target object and suppress the effect of the possible error tracking results. It is more robust than the single peak localization strategy of the original KCF tracker. Finally, this paper designs and develops a small surveillance video analysis and early warning system. The system is divided into three parts: 1) object detection module based on Faster R-CNN to identify objects of interest in input video files or stream data; 2) variable scale single filter object tracking algorithm to obtain the position and size of the target in each frame; 3) the specific abnormal behavior identification module analyzes the behavior of the target object according to the practical application requirements, and makes a warning if the specified abnormal situation is found.
【学位授予单位】：清华大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】