【作者】 刘镇弢；

【导师】 沈绪榜；

【作者基本信息】 西安电子科技大学 ， 集成电路系统设计， 2014， 博士

【摘要】 近年来，视频跟踪技术取得了显著的进展,并已在许多领域得到广泛应用。但是，目前的视频跟踪技术与实际需求仍有相当大的差距。论文从当前视频跟踪系统在实际应用中面临的问题出发，对视频跟踪系统涉及的一些关键算法及硬件实现平台进行了深入研究，重点研究了两个方面的问题：1）研究了视频跟踪系统所涉及的图像去噪声和目标跟踪算法，寻找消弱环境噪声、烟雾、遮挡等环境因素影响，提高跟踪精度的方法；2）研究视频跟踪涉及到的一些关键算法，根据算法研究分析、提取算法基本操作，建立面向视频跟踪系统的专用粗粒度的可重构硬件架构。本课题得到了国家自然科学基金项目“新一代图形处理系统芯片体系结构及关键技术研究（61136002）和“三维视频处理系统芯片动态可重构可编程体系结构研究（61272120）”的支持。作者在本文中完成的工作主要包括以下几方面：1、对图像去噪算法进行了较为系统的研究，在此基础上重点研究了基于Contourlet变换的图像去噪算法。利用非下采样Contourlet变换的平移不变性和Context模型能够实现对子带图像精细分类，并且具有较低的分类开销的特性，提出了一种结合非下采样Contourlet变换和Context模型的图像去噪算法。实验结果表明所提出的方法能够有效去除图像白噪声，与现有的一些小波域或Contourlet变换域去噪算法相比，在峰值信噪比和主观视觉效果方面均有改善。2、对视频跟踪算法进行了较为全面的研究，重点对卡尔曼滤波、扩展卡尔曼滤波、无迹卡尔曼滤波、以及标准粒子滤波等基于概率估计的跟踪算法进行了深入分析。在此基础上，为提高粒子滤波跟踪估计精度，提出一种新的结合RTS最优平滑的迭代EKF粒子滤波方法。算法主要基于RTS固定区间两步平滑算法的思路，在前向步骤中采用迭代扩展卡尔曼滤波获取非线性系统的最大后验概率估计，然后，从最后时间步开始，对滤波结果实施RTS固定区间平滑修正生成采样粒子，使其融入最新观测信息的同时更加符合系统真实状态的后验概率分布，对观测噪声具有较好的抑制能力，从而显著提高了跟踪稳定性及准确性。一维及二维模型的仿真结果验证了所提出算法的有效性。3、对视频跟踪算法中用到的图像去噪和跟踪算法进行了分析、提取了算法基本操作，并研究了适合视频跟踪应用的硬件特征。基于数据流模式和可重构技术提出了一种面向视频跟踪应用的粗粒度的可重构静态数据流硬件架构（Reconfigurable Static Data-flow hardware Architecture，RSDA），RSDA利用可重构硬件互联以及操作单元之间的硬件握手结构，克服了传统数据流实现方式的低效性，实现了架构的高度并行处理能力，并具有容易编程的特点。在Altera公司的FPGA上实现了所提结构，并与ASIC及一种优化了的GPU实现方法进行了性能对比。4、分析了所提RSDA结构的缺点，在保持所提RSDA结构的优点的前提下，采用将各种并行形式统一在一个单一结构中的思想，提出了一种更具使用灵活性和可编程性的可重构多态阵列处理器结构-PRMA，并对结构进行了初步的仿真分析。仿真结果表明PRMA能实现与RSDA相似的数据吞吐量，并有更好的可编程性。更多还原

【Abstract】 Video tracking technique has made great progress and been applied widely in manyfields in recent years. However, the difference between current technique and the actualdemand is still very large. This paper researched deeply on some key algorithms andhardware implementation platform related to a video tracking system starting from theproblems faced in practical applications, emphasizing two aspects as follows:1). thefirst is algorithms related to image denoising and object tracking, for the goal ofweakening the environmental factors such as noise, smoky fog, occlusion and so onwhich may bring an unexpected influence on tracking accuracy;2). the second is toconstruct a coarse-grained reconfigurable hardware framework oriented to videotracking applications. This work is funded by the National Natural Science FoundationProject “The Research on New Generation of Graphics Chip System Structure and KeyTechnology”(Grant No.61136002) and “The Research of Dynamically Reconfigurableand Programmable Architecture for3D-Video SoC”(Grant No.61272120).The main achievements include the following contents:1. Based on the systematic and comprehensive work on the image denoisingalgorithms, the paper put the emphasis on denoising methods based on the contourlettransformation and put forward a new image-denoising algorithm by combining thegood property of shift-invariant of Non-subsampled Contourlet Transformations withcontext model, as a result the refined classification for sub-band images is realized atlow computation cost. The experimental results prove the validity of the proposedmethod in removing the white noise, comparing with some other algorithms in waveletand contourlet domain, the PSNR and subjective visual quality of the proposed methodare both improved.2. The systematic research and analysis on video tracking algorithms is presented,including some classic probability estimation algorithms such as Kalman Filter, EKF,UKF and Particle Filter(PF). To enhance the tracking accuracy, an improved PFalgorithm is proposed based on a new proposal distribution combining iterated-EKFwith RTS optimal smoothing. Firstly, Iterated EKF is used to acquire a maximumposterior estimate of the state in the nonlinear system, and then RTS fix-intervalsmoothing is applied to correct and produce the posterior samples. The new proposaldistribution can integrate the latest observation into system and approximate the trueposterior distribution reasonably well, achieving better stability against measurementnoise and enhancing tracking stability and accuracy. Simulations of one-dimension and two-dimension models prove that the proposed algorithm is effective.3. The general image pre-processing and tracking techniques are comprehensivelyanalyzed for extracting the basic operations in the algorithms, as well as the hardwarefeatures suitable for the applications in video tracking is elaborately researched. Basedon the data-flow module and reconfigurable technique, a new framework called asReconfigurable Static Data-flow hardware Architecture (RSDA) oriented tothe applications in video tracking is proposed. RSDA can overcome the inefficiency oftraditional data-flow mode by exploiting the hardware handshake structures betweenreconfigurable hardware connection and operation units, realizing high-level parallelprocessing ability and programming easily. The proposed framework is accomplished inFPGA of Altera, compared with ASIC and an optimized GPU realization method inprocessing performance.4. The disadvantages of the proposed RSDA is analyzed. In the precondition ofkeeping the merits of RSDA architecture, the paper brought out Polymorphousreconfigurable many-core architecture (PRMA) with better flexibility in practicalapplication, and some primary analysis and simulations are presented. The simulationresults show that the PRMA can achieve similar data throughput with RSDA, and hasbetter programmability.更多还原