节点文献
可重构的无线智能视频监控平台的研究
Studies on Reconfigurable Wireless Intelligent Video Surveillance Platform
【作者】 王一木；
【作者基本信息】 浙江大学 ， 电路与系统， 2013， 博士
【摘要】 信息科技飞速发展的今天,在摩尔定律的引导下,超大规模集成电路生产技术的不断提高推动着以计算为中心的电子产业的前进。正是在这样的背景下,智能视频技术伴随着硬件能力的提升和计算机视觉科技的崛起,成为时下人们关注的热点。以智能化、网路化为特点的智能视频监控系统在安防监控、交通控制等领域应运而生。智能视频监控技术通过对前景与背景的分离,对特定区域或目标进行连续分析,从捕获的视频图像中发掘出隐藏的信息,避免了传统视频应用导致的信息爆炸。智能视频监控依赖于准确而可靠的视频分析算法,这些算法往往需要处理大量数据,消耗相当多的计算资源,给视频处理的实时性带来巨大挑战。基于DSP的视频监控系统,实现简单且具有很好的灵活性,但面对上百条视频流数据,性能往往会大打折扣；而以专用芯片作为视频分析引擎的分布式智能视频监控系统通常具有较好的实时性能,但灵活性又较差。本文综合考虑计算性能以及实现灵活两方面因素,利用FPGA的局部动态可重构技术,完成了可重构的无线智能视频监控平台的研究,主要分为以下几个方面的工作。(1)研究了Xilinx FPGA的局部动态可重构技术,提出了一种基于FPGA的可重构视频处理硬件平台,平台可在不停止工作的情况下通过无线网络连接重新对重构区域进行配置。平台中使用一种高效的ICAP控制器作为重构引擎,仅需几毫秒甚至几十微秒即可完成硬件任务的重配。(2)研究了智能视频监控的各类算法并在FPGA上一一实现,其中包括运动检测、目标识别以及目标跟踪,实现的各视频监控模块构成了可重构平台的智能视频监控IP库。其中,运动检测模块采用改进的相邻帧差算法,能够准确地标记出运动物体。目标识别模块采用改进的自映射神经网络算法,改进后的算法更加易于硬件实现,同时识别准确率与经典算法相当。目标跟踪模块采用改进的颜色粒子滤波算法,模块可在简单背景下对目标物体实现有效地跟踪。(3)研究了面向视频处理应用的高性能片上通信架构,提出了一种4元树状片上网络架构及其线上可重构算法。4元树状片上网络架构具有良好的可扩展性,可针对不同应用快速的完成配置与搭建。提出的线上可重构算法可有效地减少片内数据通信量,减轻系统负载,降低功耗。(4)提出了无线智能视频监控网络AdVision的实现,网络内支持IEEE802.11和IEEE802.15.4两种传输协议。设计了AdNode和AdBridge节点架构,并利用带有优先级的JPEG编码器,设计了低复杂度的视频压缩策略和传输机制,实现了无线网内实时、可靠地视频传输。
【Abstract】 With the rapid development of information technology, VLSI manufacturing technology, guided by Moore’s Law, is promoting electronics industry with booming computing capabilities. In this context, intelligent video technique becomes a hot issue along with the progress of computer vision technology. Intelligent video surveillance system came into being based on the application of security surveillance and traffic control. By separating foreground from background and continuous analysis in a specific area, intelligent video surveillance technology is able to get the obscure information from the captured video image and thus can avoid the explosion of information in traditional video applications. Intelligent video surveillance depends on accurate and reliable video analysis algorithms. These algorithms have to process large amounts of data and consume quite a lot of computing resources, which brings a huge challenge to real-time video processing. DSP-based video surveillance system is flexible and simple to implement, but the performance is not optimistic when processing hundreds of video streams. However, in distributed intelligent video surveillance system, ASIC chips as video analysis engine, guarantees good real-time performance, but suffers from the poor flexibility. In this thesis, considering both computing performance and flexibility, the author makes use of FPGA’s partial dynamic reconfiguration technology to carry on the study of reconfigurable wireless intelligent video surveillance platform, which can be divided into following parts.1) Xilinx FPGA partial dynamic reconfiguration technology has been studied. A FPGA-based reconfigurable video processing hardware platform which can be reconfigured through wireless connections in runtime has been proposed. The platform use an efficient ICAP controller as reconfiguration engine and only a few milliseconds or even tens of microseconds are needed to complete the process of reconfiguration.2) Various algorithms on intelligent video surveillance have been studied and implemented in FPGA, including motion detection, object recognition and object tracking. These video processing modules form the library of hardware intelligent video processing IPs. In motion detection, an improved frame differing method is used; in object recognition, a simplified self-organizing map is presented; in object tracking, an improved color-based particle filter is proposed. All these intelligent video processing IPs perform well and are implemented on FPGA.3) The high performance on-chip communication architecture for intelligent video processing has been studied. The author proposed a4-ary tree based network on chip with good scalability and flexibility which can save quite a lot of traffic load and consume less system power in communication.4) A wireless intelligent video surveillance network named AdVision has been implemented, in which both IEEE802.11and IEEE802.15.4protocols are supported. In the wireless surveillance network, AdNode and AdBridge’s hardware circuits are carefully designed to meet the requirements of both high performance and low power. To realize robust and real-time transferring, AdVision uses a low complexity compressing scheme through prioritized JPEG encoding.