群体目标识别与分析技术研究

发布时间：2018-02-13 00:47

本文关键词： 目标群航母战斗群行为分析行为识别人群场景多观察点上下文多观察点统计直方图视频描述子从属关系　出处：《华中科技大学》2016年博士论文　论文类型：学位论文

【摘要】：群体行为识别与分析是模式识别和计算机视觉领域的前沿课题,为公共场所视频监控、战场实时分析等提供有效的技术手段。随着计算机视觉技术的快速发展,基于图像的目标检测、识别等技术日趋成熟。然而,基于视频的技术仍需进一步提高,特别是行为识别。群体行为识别是行为识别的一种,其场景比一般行为复杂。到目前为止,不同的目标群体很难利用固定的算法进行分析。相对单个目标或多个目标,群体行为是整体行为,目标与目标之间的上下文关系密切。通常具有以下特点:目标数量较多,运动环境复杂,速度快慢不一,密度高或相互之间遮挡严重等。目前,国内外学者对该课题研究不多。虽然近几年取得了一些阶段性的成果,出现了一些大型数据库,但是不同课题组研究的切入点不同,整体研究处于比较分散的阶段。本文的研究主要集中在两类目标群。一是稀疏目标群(例如航母战斗群);二是稠密目标群(例如人群)。本文以航母战斗群为例研究了稀疏目标群;以人群为例研究了稠密目标群。针对群体目标的特点,本文的研究内容如下:首先,模拟了卫星监视中的航母战斗群航行视频。航母战斗群的数据非常宝贵,模拟航母战斗群队形变化航行的意义重大。虽然目前的卫星技术很难支持大范围的视频拍摄,但是模拟视频可以验证识别算法在任意时刻的有效性。为了尽可能真实模拟航母战斗群的航行,本文分析了雷达侦察卫星和光学侦察卫星监视航母战斗群的可行性。提出利用三次Hermite插值函数规划军舰的轨迹,既可以保证规划的轨迹函数二次可导,也能很好控制队形变化过程中舰船之间的距离,防止发生碰船事件。最后,为了增强模拟航行的视频真实性,采用“谷歌地球”中的航母和军舰照片作为军舰模板,动态海面作为背景,沿着设置好的轨迹生成航母战斗群的模拟航行视频。其次,在假设舰船目标已经检测出来的基础上研究了航母战斗群的队形识别和行为分析。提出在阿基米德螺线上选取一系列观察点,计算每个观察点与航母战斗群的上下文信息,形成了多观察点上下文描述子,成功解决了旋转和尺度不变性问题。建立了概率密度函数模型,将队形的局部信息与全局信息有效融合,增强了算子的描述能力。该描述子的维度与军舰数量无关,其识别性能对航母战斗群的中心区域军舰数量不敏感,符合航母战斗群编队的实际情况。提出了基于隐马尔科夫模型的行为识别方法,并在不同的模拟视频中验证了算法的有效性。针对人群检测,本文提出了一种新的局部区域描述子——多观察点统计直方图。在多个观察点上进行径向梯度变换,形成了一种局部区域的整体描述子。在不需要归一化图像尺度大小的情况下,能统一描述不同尺度的人群图像块。最后结合了快速目标框的方法进行人群检测。基于目标框的方法人群检测不需要高斯金字塔,但却可以利用少数目标框覆盖大部分目标。针对形态变化较大的人群场景,提出了一种多标签的分类器模型。人群行为种类繁多,类别之间关系复杂,传统的分类器效率不高、效果不好。充分利用不同类别之间的从属关系,提出了一种高效分类器模型。相比传统分离器,本文的分类器具有完美的闭合解,可以同时高效处理多个类别。人群场景分类是一个多实例的问题,本文结合了深度卷积网络和Fisher Vector(FV)编码,构建了具有时空信息的视频描述子,高效处理了多实例问题。以上方法均在充足的数据集和实验下进行了验证。实验结果表明,本文提出方法大部分结果优于主流方法。大部分数据集是实际拍摄的视频数据,提出的方法具有很强的实际应用价值。本文的方法可以视为一些基本的技术,具有相关领域潜在的实际应用价值,例如,事件检测、行为识别等。
[Abstract]:Behavior recognition and analysis of the group is a leading research field of pattern recognition and computer vision, video surveillance for public places, provide effective technical means real-time battlefield analysis. With the rapid development of computer vision technology, image target detection based on recognition technology is becoming more and more mature. However, the video technology needs to be further improved based on the special is the behavior recognition. Group behavior recognition is a kind of behavior recognition, the scene is more complex than the general behavior. So far, different target groups are difficult to analyze using the fixed algorithm. Compared with single target or multiple targets, group behavior is the overall behavior context between the target and the target is usually close. The following characteristics: a large number of target motion in complex environment, the speed of a high density or mutual occlusion seriously. At present, domestic and foreign scholars on the subject. There is not much. Although in recent years has made some achievements, there are some large databases, but the starting point of different research group, the overall research in a relatively dispersed phase. This study focused on two types of target groups. One is the sparse target groups (e.g. carrier battle group); two is the dense target group (e.g. population). The aircraft carrier battle group as an example to study the sparse target group; population as an example to study the dense target group. According to the characteristics of the target groups, the research contents of this paper are as follows: firstly, the simulation of satellite surveillance of aircraft carrier battle groups sailing video carrier battle group data. Very valuable, simulation of the aircraft carrier battle group formation changes sailing of great significance. Although the satellite technology is difficult to support a wide range of video capture, but analog video recognition algorithm can be verified at any time for effectiveness. As far as possible to simulate navigation aircraft carrier battle group, this paper analyzes the radar reconnaissance satellite and optical reconnaissance satellite surveillance aircraft carrier battle group. The feasibility of proposed using three Hermite interpolation function planning ship trajectory, which can not only ensure the planning path function can guide two times, can well control the ship formation changes in the process of distance the ship, to prevent the occurrence of touch events. Finally, in order to enhance the authenticity of the video simulation of navigation, the "Google earth" in the aircraft and warships warship photos as template, dynamic sea as background, set up along the trajectory generation of aircraft carrier battle groups sailing simulation video. Secondly, based on the ship target detection has been assumed on the study on the analysis of formation recognition and behavior. The carrier battle group selected a series of observation points in Archimedes spiral, calculated for each observation point and the aircraft carrier battle group The context information, the formation of multi observation point context descriptor is solved successfully, rotation and scale invariance. A probability density function model, local information and global information formation and effective integration, enhance the operator description ability. The number of dimensions and warships the descriptor is independent of the recognition performance is not sensitive to the number of regional center Navy aircraft carrier battle group, in accordance with the actual situation of the aircraft carrier battle group formation. Put forward the behavior recognition method based on Hidden Markov model, and verify the effectiveness of the algorithm in the analog video. According to different crowd detection, this paper proposes a new local descriptor, multiple observation points of radial histogram. The gradient transform in a plurality of observation points, forming a whole a local descriptor. Without the need of normalized image size under the condition of uniform People describe image blocks at different scales. Finally the method of fast target frame for crowd detection. Detection method of target population box does not need Gauss in Pyramid based on, but can use a box cover most of the target object. According to the morphological changes of large crowd scenes, proposed a multi label classifier model. The crowd behavior types there are categories of the relationship between the complexity of the traditional classifier, the efficiency is not high, the effect is not good. Make full use of the dependencies between different categories, this paper presents an efficient classifier model. Compared with the traditional separator, the classifier has the perfect solution, at the same time, can handle multiple categories. The crowd scene classification is a multi examples of problems, combining the convolutional network and Fisher Vector (FV) encoding, construct the video descriptors with spatial and temporal information, efficient handling of multiple instances All the above problems. The method was validated in sufficient data sets and experiments. The experimental results show that this method is better than most of the results of mainstream method. Most of the data set is the actual shooting video data, the proposed method has strong practical application value. This method can be regarded as some of the basic technology, practical related areas of potential application value, for example, event detection, behavior recognition.

【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】