【作者】 潘翔；

【导师】 顾伟康； 叶秀清；

【作者基本信息】 浙江大学 ， 通信与信息系统， 2003， 博士

【摘要】 运动目标检测，既古老——历史悠久，又新颖——研究方法日新月异。面对不同的研究对象，运动目标检测既有一般指导意义的理论研究，又有工程背景的特殊性研究。本文注重在理论指导下的特殊性研究，即在联合利用时间信息、空间信息、频率信息以及目标特征能够提高运动目标检测能力的思想指导下，分别研究了声纳的运动目标检测和基于计算机视觉的交通车辆检测。 随着降噪技术的发展，安静型目标的出现，声纳技术面临低信噪比的挑战，武器射程的增加又对声纳提出了更高的要求，关系到声纳的四大功能(检测、定位、识别和跟踪/目标运动分析)第一和第四个的运动目标检测任务重大，因为它对其余两个起着基础性的指导作用。它不仅与信道(环境)知识有关而且与目标的特性密切相关。 本文通过了对模基信号处理实现信号增强的理论研究，提出了一种利用传播模型、噪声模型和阵测量模型、并注重环境信息来实现水声信号增强的方法，继而实现了基于增广高斯—马尔可夫过程和相应的扩展卡尔曼滤波联合的模基辨识器的信号增强算法。实验数据和模拟数据的结合验证了模基辨识器不仅能够产生增强的水听器阵处的声压场表示，而且能够输出模域表示(模深度函数和水平波数)和目标的方位(平面波模型)，也具有较好的自适应性及对失配的宽容性。 在后处理中，提出了利用特征增强提高目标探测能力的思想，研究了用信号微结构表征目标特征的方法。基于信号微结构知识，分析了水声信号的时-频特性和分布特性。为了加强对弱信号的检测和跟踪，在特征增强的基础上，研究了特征提取的方法，开发了双阈值Viterbi线谱跟踪器。模拟数据的结果表明信号微结构的表征方法是正确的，特征提取是有效的。 将信号微结构的提取重要环节，利用TMS320C40 DSP芯片的高速和并行处理能力，完成了算法向硬件平台的映射，构成了实时的目标探测系统。在湖上试验中，该系统能稳定地跟踪水声目标，这说明信号微结构的提取方法是有效的，适合工程化应用。 随着公路交通的快速发展，现有的交通视频检测系统显得力不从心，不仅视频图像信息的获取和处理的速度较慢，而且车辆检测的准确性也低，前者依赖于硬件浙江大学博士学位论文设备性能的提高而改善，后者则需要研究和开发更先进的车辆检测算法。本文主要研究车辆检测的关键算法和设计实时的车辆检测系统。 现有的车辆检测系统之所以出现漏检和误检的原因在于:l)灰度图用于检测的信息少;2)背景模型不能反映变化的道路背景;3)存在阴影。为此，本文在研究彩色模型的基础上，提出了快速车辆的检测方法，继而研究和开发了两种背景模型，即，统计背景模型和确定的自适应背景模型。根据阴影形成的机理及其特性，本文提出了两种阴影检测方法，即，蓝波段信息方法和确定的非模型化方法。实际数据处理的结果表明利用背景模型能够有效地检测快速运动目标，而且两种阴影检测方法都能够准确地检测阴影。 构建了基于彩色分割和阴影抑制的实时的车辆检测系统。在降噪预处理的基础上，联合背景抑制和边缘提取进行运动目标检测，通过阴影抑制进一步提高了车辆检测的准确性。经过实际数据的验证该系统能够比较好地对快速行进的车辆进行检测和跟踪，并且能够正确地检测阴影。此外，该系统还适合一般的公路交通或市区交通的车辆检测。 从表面上看，基于声学方法的水面或水下的运动目标检测与用光学的方法检测高速公路或一般道路上的运动目标大相径庭。然而从信号处理的观点来看，二者有许多相似之处，特别是分别通过声电转换和光电转换之后，运动目标的信号微结构表征和抽取技术与图像处理技术。本论文将两个领域的有关工作统一在图像信号处理的运动目标检测框架下以期产生交叉，共同促进动目标检测技术的发展。更多还原

【Abstract】 Moving target detection has both a long history and novel approaches. In face of different objects, there are both theoretical study for a guide and application with particularities. Baesd on the theory that moving target detectability can be enhanced by joint space-time processing with the application of frequency features and targets’ characteristics, we focus our attention on particularies of sonars’ target detection using underwater acoustic techniques and vehicle detection using computer vision techniques.With the development of reducing noise technology and the advent of quieter targets, sonar faces up to the challenge of low signal-to-noise ratio detection while high detection performance is demanded by weapons’s increasing range.Moving target detection is two of first and last of sonar’ four main tasks: detection, loacalization, classification and tracking/motion analysis of targets, which is the base of other two functions. It is not only tightly related to transimission channel (environment) information but also the characteristics of targets.In the study of environment model-based processing,the thesis presents a signal enhancement method based on combination of the ocean acoustic propagation, measurement system and noise models.Furthermore,a signal enhancement algorithm is developed on the basis of the model-based identifer (MBID) implemented by the the augmented Gauss-Markov process and corrsponding extended Kalman filter. The experiment shows MBID can produce the enhanced pressure field at hydrophone array, provide modal domain representation of pressure (modal functions and horizontal wavenumbers) and target’s bearing (plane waves), and have good adaptive ability and robustness against mismatch.In post processing after signal enhancement, the feature enhancement using signal micro-structure characterizing target features is exploited to improve target detectability. In view of the signal micro-structure knowledge, more attention is paid on analyzing the time-frequency and distribution characteristics of underwater acoustic signals. To detect and track weak signals, feature extraction is studied on the base of feature anlysis, and a Viterbi spectrum line tracker with double thresholds is exploited. The simulation shows that feature enhancement and extraction based on the signal micro-structure characterization are efficient.The algorinthms of the signal micro-structure extraction as a key part of feature enhancement are mapped into hardware platform to build a real-time target detection system by developing high speed and parallel processing ability of DSP chips of TMS320C40. That the system can stably track targets at lake test shows the signal micro-structure extraction is effective and can be implemented for engineering application.In contrast to the development of freeway, the existing intelligent transportation system (ITS) lags behind for its image acquiring and processing with low efficiency and low acuracy vehicle detection. The improvement of ITS depends not only on the development of hardware platform, but also fast and reliable algorithms for detectingvehicles. The thesis is to exploit key vehicle detection agorithms, and design a real-time vehicle detection system through the detection algorithm mapping.The exiting ITS has a low detection and high false alarm rates due to:l)less information of gray images for detection;2) lack background model adaptive to real background change; 3)existence of shadows. In color space, the thesis proposes a detection method for high speed vehicles with two background models, one statistical background model and another deterministic adaptive background model. After the underlying physics of shadows and their characteristics are studied, two shadow detection approaches are exploited, one blue wave band information approach and another deterministic non-modeled approach. The experimnet shows that it is effficient to detect high speed vehicles by means of background models and shadows can be correctly detected with shadow detection algo更多还原