【作者】 李宏友；

【导师】 汪同庆；

【作者基本信息】 重庆大学 ， 仪器科学与技术， 2009， 博士

【摘要】 基于视频的目标检测与跟踪是计算机视觉领域主要的研究方向之一。它在智能监控、人机交互、视觉导航等众多领域有着广泛应用,并发挥着举足轻重的作用。本文围绕视频目标检测与跟踪技术,重点研究了运动检测、目标匹配、跟踪框架等关键技术及方法。具体地,论文的主要工作和贡献集中在以下方面:①提出了3D OGHM (Orthogonal Gaussian-Hermite Moments)运动检测方法。根据视频序列图像特有的3D特性,本文从整个视频空间的角度来思考运动检测问题。在考虑1D时间信息的同时,加入了2D空间信息用于运动检测,提出了一种联合时空信息的3D OGHM运动检测算法。3D OGHM一方面继承了时差运动检测法的快速性优点。另一方面,它充分发挥了视频信息特有的3D特性,具有较强的运动提取能力。此外,该算法还具有显著的抗干扰能力。②提出了一种带距离约束项的基于亮度信息的主动轮廓模型IMDC (Intensity-based Model with Distance Constraint)。针对传统几何主动轮廓模型需要重新初始化或需要借助原始图像的梯度信息来保证水平集进化如期收敛这一问题,论文提出了一种不同版本的(几何)主动轮廓模型IMDC。IMDC模型引入一个距离约束项作为内部能量来确保水平集函数始终不偏离符号距离函数,避免了进化过程中对水平集函数的不断初始化。同时,借鉴C-V模型的基本思想,IMDC采用图像的亮度信息而非梯度来构造模型的外部能量项,确保了零水平集曲线稳定地收敛于期望的图像特征点(如目标轮廓点)。实验结果表明,IMDC模型不仅有效克服了传统模型需重新初始化或无法应对弱边缘特征等问题,而且具备全局的分割能力和较强的抗噪性能;水平集的初始化操作也非常简单、灵活。③提出了ADC(Active Drift Correction)仿射配准算法。针对传统图像仿射配准算法在算法效率和鲁棒性方面的问题,本文提出了一种新的解决思路:采用合成目标能量代替传统的单一能量成分作为仿射配准算法的目标表达式,即“合成目标能量”思想。基于该思想,文章提出了主动漂移矫正算法(ADC)。ADC算法通过引入一个漂移矫正项,来改进传统仿射配准算法的目标能量函数,增强了算法抗漂移的能力,提高了算法的鲁棒性。改进后的算法不需要传统算法中为增强鲁棒性而采用的许多复杂措施(如分块操作或二次跟踪),因此算法在效率、鲁棒性等方面具有较高的综合性能。此外,文章对仿射配准跟踪的更新环节也进行了深入地分析和优化改进。④研究了适用于动静态背景及多种应用的视频目标检测与跟踪混合框架。通过整合本文提出(或改进)的各种独立算法,研究了一套基于视频的目标检测与跟踪混合框架。该混合框架以分别侧重于检测和匹配的两套子框架为基础,适用于固定或动态场景下多个或特定目标的检测与跟踪应用。特别地,针对固定背景下以检测为中心的跟踪子框架,在运动检测阶段,文章采用了局部OGHM运动检测法来代替全局的3D OGHM检测,明确了“潜在目标邻域”的基本概念及其确定方法,进一步提高了运动检测的效率;在目标分类识别阶段,文章提出采用BM(Bidirectional Map)双向映射算法,有效解决了多个混杂目标之间的匹配问题。通过对该混合框架的研究和试验,进一步证实了本文提出的各种算法的有效性。更多还原

【Abstract】 Video-based target detection and tracking is one of the research hotspots in the field of computer vision. It plays a very important role in many applications, such as smart surveillance, human-machine interface and visual navigation. This thesis aims at researching the technologies of target detection and tracking, which focuses on the key issues on motion detection, matching and framework of tracking.The main contributions of this dissertation are listed as follows:①A new method named 3D OGHM (Orthogonal Gaussian-Hermite Moments) for motion detection is proposed. This thesis considers the motion detection at the point of all dimensions of the video space. By combining the temporal and the special characteristics of the video sequences, the 3D OGHM algorithm is proposed. As one of TVA (Temporal Variation Analysis) methods for motion detection, it has high efficiency firstly. Additionally, because the 3D property of video is considered, it has a stronger ability to enhance the motion information in comparison with the congener methods. Besides, it is outstanding in anti-jamming as well.②A new version of geometric active contour modal (ACM) named IMDC (Intensity-based Model with Distance Constraint) is proposed. Most of current ACMs either have to re-initialize the level set function constantly or require the gradient flow to stop the evolution of the curve. To solve this problem, the internal energy term of IMDC imports a distance constraint that penalizes the deviation of the level set function from a signed distance function (SDF). And the external energy term of IMDC adopts the intensity instead of the gradient of the image to drive the curve on zero level set toward the desired image features, such as the object boundaries. The experimental results show that the model presented efficiently avoids the re-initialization and overcomes the problem that the traditional models can not work well with the images with low gradient. Moreover, our model is able to acquire the global optimization of the segmentation and it has a good anti-noise performance. The initialization is also simple and flexible as well.③An improved affine image alignment (AIA) algorithm called ADC (Active Drift Correction) is proposed. To solve the traditional problem of worse compatibility between the robustness and the efficiency of AIA. The basic idea of ADC is to incorporate a drift correction term into the traditional goal energy function, which also called“component goal energy thought”(CGET). Based on CGET, ADC has the ability of anti-drift, which boosts its robustness. Moreover, many extra techniques (e.g. dividing blocks or second tracking) in traditional methods for a high robustness are unnecessary. The experimental results show that our algorithm is simple and efficient. It achieves a higher performance of robustness than the traditional methods. However, it makes no compromise with the complexity and real time performance of the algorithm. Besides, the dissertation analyzes and researches the update strategies of AIA.④A mixed framework (MF) for target detection and tracking is introduced. By combining the independent algorithms proposed in this thesis, a mixed framework (MF) for video-based target detection and tracking is introduced. This framework which includes two sub-frameworks emphasizing respectively on the detection and the match can be used in the case of fixed or moving background. It also fits these cases that multiple targets or special (known) target should be detected or tracked. Specially, in the DE (Detection Emphasized) sub-framework, local OGHM instead of global 3D OGHM is adopted for higher detection efficiency. And a new method named BM (Bidirectional Map) is introduced to resolve the target recognition and matching problems. The results of the MF research and experiments futher show that the algorithms proposed in this thesis are effective.更多还原