【作者】 张超；

【导师】 吴小培；

【作者基本信息】 安徽大学 ， 计算机应用技术， 2014， 博士

【摘要】 在现今复杂的安全形式和迫切的市场需求共同推动下,智能视频技术作为一种全新的安全监管手段,已成为近年来信息科学和安防技术领域中的研究和应用热点。智能视频将传统视频监控技术和现代模式识别图形图像处理算法结合,在无需外界干预的情况下,通过对视频数据流的分析,自动完成场景中目标的检测和跟踪,并在此基础上进行目标行为理解和辅助决策,实现人不在回路中的智能监控,带来了监控技术的革命。尽管智能视频的基本流程很早就被提出,相关核心算法也已经过了较长时间的发展,但智能视频技术在实际应用中仍然面临一些困难：现有算法易受真实环境中光照变化等各种干扰因素的影响,目标的特殊运动状态以及与背景环境区分性不佳同样会导致算法性能的急剧下降。因此,能有效应对复杂背景条件和目标特殊运动状态、准确高效且稳定性好的运动目标检测与跟踪算法仍将是很长一段时间内研究者们追求的目标。本文立足于智能视频分析技术的应用研究,对智能视频系统中的运动目标检测和目标跟踪等若干核心关键问题进行研究,全文涉及到的主要工作包括以下几方面：1.针对现有传统GMM算法应对低速目标时易出现前景破碎的性能局限,提出了基于前景模型匹配和短时稳定度的运动目标检测算法。算法对传统GMM算法中背景模型匹配失败时生成的前景模型加以利用并引入前景模型匹配和更新机制防止前景模型向背景模型转换导致的慢速目标漏检,同时提出短时稳定度指标衡量一段时间内的像素值波动以应对低区分性目标。实验结果表明短时稳定度和前景模型结合增强了算法对低速低区分性目标的适应性。2.提出了一种基于前景模型匹配和短时稳定度的遗留物检测算法。遗留物是慢速目标的极限情况,算法中使用前景模型表征遗留物,并以最高的优先级保证其先于背景模型和后来像素值进行匹配,降低了背景模型与遗留物误匹配的错误风险,防止遗留物被背景吸收。多场景下的遗留物检测实验结果验证了算法的可行性。3.在多维统计信号分析理论框架下提出了基于十六通道独立分量分析算法并综合四种不同观测信号生成方式的运动目标检测算法。独立分量分析和主分量分析算法是现今较有代表性的多维统计信号分析方法,现有基于独立分量分析和主分量分析的运动目标检测算法大多使用单一观测信号生成方式和双通道信号进行检测,无法为前景分离提供更多有效信息,常导致目标检测不完整。大量的对比实验验证了不同通道数和观测信号生成方式下ICA算法的检测性能差异,证明了使用大通道数和多种观测信号生成方式带来的检测性能提升,同时也验证了多维统计方法和其他多种运动目标检测方法在算法特性上存在的差异。4.针对现有基于Meanshift的目标跟踪算法使用固定量化阶数生成核直方图的问题,提出了使用直方图动态量化级的Meanshift目标跟踪算法。算法在大景深场景中根据目标尺寸的变化动态改变直方图阶数,并使用动态时间规整算法进行不同维度特征的匹配。对比实验结果表明改进算法能在大景深场景中自动调节量化级阶数,并以此达到更低的平均帧处理耗时。5.作为对前述理论的综合和工程化运用,在上述智能视频核心算法的基础上设计实现了一套智能视频系统。按照智能视频的一般流程分别给出了各个模块的开发过程并在此基础上集成为一套具备智能监控功能的实用化系统。系统可以实现对多通道视频数据的实时智能监控,具有虚拟警戒区域设置、入侵报警、关键帧保存、入侵视频回放和环境参数集成显示等功能,对智能视频系统的具体功能和应用价值进行了更直观的演示验证。更多还原

【Abstract】 Motivated by today’s complex security situation and urgent application requirements, intelligent video technology, as a new method for security supervision, has become a hotspot in both scientific research and engineering technology. Intelligent video surveillance technology combines traditional video surveillance techniques with modern pattern recognition and graphics/image processing algorithms, it automatically detects and tracks objects by analyzing the video stream without any manual intervention. Even more, behavior understanding and assistant decision-making can be achieved based on the objects’information obtained from detection/tracking module. Intelligent video surveillance technology brings a Man-Not-in-the-Loop automatic monitoring, which leads to a revolution in video surveillance technology.Although the standard flow of intelligent video has been proposed and the core algorithm has also been developed for a long time, the application of intelligent video technology still faces some difficulties:current algorithms are vulnerable to a variety of interfering factors such as illumination changing in real environment, the object’s special moving pattern as well as its indistinguishable surface may also lead to a sharp decline in performance. Therefore, researchers are still looking forward to more effective and stable algorithms which are capable of precisely dealing with complex background and special motion patterns in object detection and tracking. In this dissertation, core algorithms in intelligent video system such as moving objects detection and tracking are researched, the research work in this dissertation mainly includes the following aspects:1. Aiming at the problem that existing GMM algorithm prone to get broken target when dealing with slowly moving object, a moving object detection algorithm based on foreground model matching and short-term stability measure was proposed. Potential foreground models that generated at each pixel were used to match the incoming pixel and a matching mechanism was developed to update the foreground models. Meanwhile, the short-term stability measure was employed to deal with foreground component which was fluctuating in its neighboring region. Experimental results showed that the combination of foreground matching and short-term stability measure enhanced the adaptability to slowly moving objects.2. An abandoned object detection algorithm based on foreground model matching and short-term stability measure was proposed. Abandoned object is the extreme case of slowly moving object. The proposed algorithm used foreground models to characterize the abandoned object, and ensure their the highest priority to firstly match the following pixel values rather than background model s, reducing the risk of foreground pixel mismatching the background models and hence prevent the abandoned objects being absorbed into background. Abandoned object detection results under multiple scenarios verified the feasibility of the proposed algorithm.3. A sixteen-channel ICA based motion detection algorithm using four observation vector generation methods was proposed. Independent component analysis and principal component analysis were typical multi-dimensional statistical signal analysis methods. The existing moving object detection algorithm based on independent component analysis and principal component analysis using just single method for observation vector generation and two-channel data for separation, unable to provide more effective information for separation, and hence resulting in incomplete detection result. A large number of comparative experiments shown the difference on detection performance under different observation signal generation methods and different channel numbers, and validated the improvement by selecting larger channel number and using four different observation signal generation methods. The characteristic differences between statistical methods and a variety of other traditional methods are also verified in the experiments.4. Aiming at existing Meanshift based object tracking algorithm using fixed quantization order to generate a histogram, an improved Meanshift object tracking algorithm using adaptive quantization order in color space was proposed. Quantization order was dynamically changed according to the size changing of the moving objects, and DTW was employed to measure the similarity of features with different length. Comparative experiments demonstrated that the improved algorithm improved the calculation flexibility and achieved a lower average per-frame time consuming.5. As an engineering application of above theories, an intelligent video surveillance system was designed and implemented based on a set of core intelligent video algorithms. In accordance with the general flow of intelligent video processing, the detailed development process of each module were given and an intelligent video surveillance system which can be put into practical use was integrated based on the individual modules. The system realized many real-time multi-channel intelligent monitoring functions such as virtual alert line setting, intrusion alarm, key frame preservation, invasive video playback and integrated environmental parameters displaying. It well demonstrated the specific functions and application value of intelligent video surveillance system.更多还原