基于DSP优化的行人识别算法在智能监控中的研究与应用

发布时间：2018-03-17 10:45

本文选题：C64x+　切入点：智能监控系统　出处：《电子科技大学》2014年硕士论文　论文类型：学位论文

【摘要】：近年来,随着技术的飞速发展,视频监控系统从第三代网络监控逐步过渡到第四代智能监控。在智能监控中,视频/图像处理已经成为重要研究课题,运动目标检测,特别是行人检测已经成为智能监控重点研究内容。本文研究了视频处理中运动目标提取和行人检测算法,设计原型系统在全高清智能监控系统(1080P)上完成运动行人检测。本文的主要工作有:1.对整个智能监控系统的功能进行分析,针对传统智能监控架构的缺点,采用DSP智能相机设计了一种分布式的智能监控架构——利用DSP完成运动目标提取,在运动区域上使用PC完成行人检测。2.仔细分析现有行人检测算法,针对传统多尺度滑窗检测速度慢,效率低的缺点,根据应用场景,应用行人总是在运动这个先验信息,设计算法对视频中的运动区域进行检测,有效降低了检测窗口的数目。3.根据DSP处理器的特性,分析对比多种运动检测算法,最终选择Vibe进行运动检测。在运动检测的基础上,设计了一种基于运动区域的行人扫描方法,代替传统的多尺度滑窗扫描。4.详细分析HOG、HOGLBP特征在行人检测中的优势,针对这两种特征在分类时计算量大的问题,提出采用主成分分析(PCA)对HOGLBP特征进行降维。通过理论分析和实际验证证明了PCA_HOGLBP特征在保证行人检测效果的同时,能有效减低分类的复杂度。5.根据智能监控系统的总体设计和运动行人检测算法,实现了该原型系统。仔细分析了该系统算法的耗时问题,使用编译器选项和双缓冲操作对代码进行优化。在认真分析算法结构的基础上,使用Intel扩展指令集SSE对向量内积运算进行了针对性优化,提出使用C64x+指令对3×3中值滤波算法进行优化。实验表明,在同等条件下,本文优化的3×3中值滤波算法比TI在imgLib库中提供的函数快5.24倍。本文设计的智能监控系统将运动检测算法被分散到各个DSP智能相机,PC只对疑似运动区域进行扫描和分类,占用系统资源低;只在异常情况下传输高码率视频流,网络负载低。实验证明,本文设计的智能监控系统在全高清视频下能有效的对运动行人进行检测,检测速度快,检测效果好,漏检和误检少。
[Abstract]:In recent years, with the rapid development of technology, video surveillance system from the third generation network monitoring gradually transition to the fourth generation of intelligent monitoring. In intelligent monitoring, video / image processing has become an important research topic, moving target detection, especially pedestrian detection has become the key research contents of intelligent monitoring. This paper studies the moving object in video processing extraction and pedestrian detection algorithm, the design of the prototype system in full HD intelligent monitoring system (1080P) on moving pedestrian detection. The main works of this paper are: 1. the function of intelligent monitoring system are analyzed, the traditional intelligent monitoring architecture using DSP smart camera design flaws, the architecture of a distributed intelligent monitoring using DSP complete moving object extraction, the use of PC in the motion area complete pedestrian detection.2. careful analysis of the existing pedestrian detection algorithm, based on the traditional multi-scale Sliding window detection speed, low efficiency, according to the application scenarios, pedestrians always in motion this prior information, the design algorithm to detect the motion region in video, effectively reducing the detection window number.3. according to the characteristics of DSP processor, analysis of a variety of motion contrast detection algorithm, the final choice of Vibe in motion detection. The basis of motion detection, design a pedestrian scanning method based on motion region, multi-scale sliding window scanning.4. HOG detailed analysis to replace the traditional HOGLBP feature, advantage in pedestrian detection, according to the two characteristics of the large amount of calculation in the classification, proposed using principal component analysis (PCA) to reduce the the dimension of the HOGLBP feature. Through theoretical analysis and practical verification proved that the PCA_HOGLBP feature in ensuring the detection effect of pedestrians at the same time, can effectively reduce the complexity of.5. classification based on intelligent monitoring system The overall design and the pedestrian detection algorithm, realized the prototype system. Careful analysis of the time-consuming problem of the algorithm of the system, optimizes the code compiler options and double buffer operation. Based on analyzing the structure of the algorithm, using the Intel instruction set extension to SSE amount of inner product operation of the targeted optimization, put forward to 3 x 3 median filtering algorithm was optimized using the C64x+ instruction. Experimental results show that under the same conditions, the optimization of the 3 x 3 median filtering algorithm is 5.24 times faster than the TI function in the imgLib library provides the intelligent monitoring system designed in this paper. The motion detection algorithm is distributed to every DSP smart camera, PC scan and the classification of suspected motion regions, low system resource use; transmission of high bit rate video stream only in exceptional circumstances, the network load is low. The experiment proves that the design of the intelligent monitoring system in full HD video It can detect the moving pedestrians effectively, the detection speed is fast, the detection effect is good, the leak detection and error detection are few.

【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.41;TN948.6

【相似文献】