基于DM8127的行人检测智能前端设计与实现

发布时间：2018-03-28 02:04

本文选题：智能前端　切入点：行人检测　出处：《大连海事大学》2017年硕士论文

【摘要】：随着政治、经济的发展,各个国家、企业、个人越来越关注安防事业。监控系统由最原始的模拟视频和人眼监测到中期的半数字化存储再到如今的全数字化监控系统,互联网的发展、编解码算法的升级都功不可没。而智能前端监控系统在监控系统中脱颖而出,算法的多样性需求和前端处理器的飞跃发展,使得智能前端监控系统的广泛应用成为必然。多年来,行人检测课题的研究持续不断。行人检测算法在电子卡口、无人车行人避让系统、客流量检测等应用中作为基础算法有着至关重要的作用。行人检测智能前端是带有行人检测分析功能的智能前端,不但能够代替人眼进行监查,而且能够减少传输信号所占用的带宽和存储资源。本文根据《安防监控视频实时智能分析设备技术要求》设计了行人检测智能前端系统的功能和性能要求。通过分析DM8127的优势,确定以DM8127为主处理器的网络摄像机作为系统实现的硬件平台,并分析了行人检测智能前端的五个模块,同时结合系统的软硬件平台,选用支持向量机(Support Vector Machine,SVM)和梯度方向直方图(Histogram of Oriented Gradient,HOG)相结合的方法作为前端分析模块的实现方案。在MATLAB上模拟了行人检测系统,包括提取HOG特征模块、图像金字塔检测模块以及多窗口融合模块。根据智能前端行人检测的实时性和准确率要求,针对HOG特征的三个缺点给出相应的解决方法:1)HOG特征的缩放不变性差。选择包含不同高度行人的同尺寸图像数据集,通过图像金字塔检测原理,设计三层图像金字塔,并在各层进行行人检测,分析HOG特征缩放不变性性能,结论为64*128大小的检测窗口可以检测到行人高度范围在88像素-128像素内的行人。根据此结论给出了单层图像金字塔检测法。2)HOG特征的特征维度较高。特征维度过高导致提取特征耗时长,检测速度缓慢,可在保证准确率的前提下进行算法参数的适当调整。3)HOG特征对被遮挡的行人,检测效果较差。挑选只有上半身包含头肩信息的行人作为部分训练正样本。检测到行人后,根据坐标在界面显示模块将行人框出,在多尺寸窗口融合技术的原理上,给出了递归的窗口融合算法。仿真后,选用LIBSVM移植到DM8127上用于行人判别,并移植提取HOG特征的算法,实现整个智能监控系统并进行测试。实测结果表明:添加只包含头肩信息的行人做为训练正样本,可以有效地解决行人下半身被遮挡的问题;通过调整HOG特征的提取参数,可以在保证精度符合要求的情况下,有效提高检测速度;给出的递归的窗口融合算法,可以有效地将多个窗口融合;HOG和SVM相结合的算法移植到DM8127中可以检测到90%以上的行人。
[Abstract]:With the development of politics and economy, various countries, enterprises and individuals are paying more and more attention to the security cause. The monitoring system is from the most primitive analog video and the human eye to the semi-digital storage in the medium term and then to the full-digital monitoring system now. The development of the Internet, the upgrading of coding and decoding algorithms, and intelligent front-end monitoring system stand out in the monitoring system, the diversity of algorithms and the rapid development of front-end processors, It makes the wide application of intelligent front-end monitoring system inevitable. Over the years, the research on pedestrian detection has continued. Pedestrian detection algorithm in electronic bayonet, unmanned pedestrian avoidance system, The intelligent front end of pedestrian detection is an intelligent front end with the function of pedestrian detection and analysis. Moreover, it can reduce the bandwidth and storage resources of transmission signal. According to the Technical requirements of Real-time Intelligent Analysis equipment for Security Surveillance Video, this paper designs the function and performance requirements of intelligent front-end system for pedestrian detection. By analyzing the advantages of DM8127, this paper analyzes the advantages of this system. The network camera with DM8127 as the main processor is determined as the hardware platform of the system, and the five modules of the intelligent front end of pedestrian detection are analyzed. At the same time, the hardware and software platform of the system is combined with the hardware and software platform of the system. Support vector machine (SVM) and gradient histogram of Oriented histogram (histogram of Oriented gradient histogram) are selected as the implementation of the front-end analysis module. The pedestrian detection system is simulated on MATLAB, including extracting the HOG feature module. Image pyramid detection module and multi-window fusion module. According to the real-time and accuracy requirements of intelligent front-end pedestrian detection, Aiming at the three shortcomings of HOG feature, this paper gives the corresponding solution, that is, the scaling invariance difference of the HOG feature. The same size image data set including different height pedestrians is selected, and the three-layer image pyramid is designed by the principle of image pyramid detection. At the same time, pedestrian detection is carried out in each layer, and the scaling invariance of HOG features is analyzed. Conclusion the detection window with the size of 64m 128 can detect pedestrians whose height ranges from 88 pixels to 128 pixels. According to this conclusion, the feature dimension of pyramid detection method of single-layer image is higher and the characteristic dimension is too high. It takes a long time to extract the feature, The detection speed is slow, and the algorithm parameters can be adjusted properly on the premise of ensuring the accuracy. The detection effect is poor. Only the upper half of the pedestrian with head-shoulder information as part of the training positive sample. After detecting the pedestrian, according to the coordinates display module in the interface to frame the pedestrian, in the principle of multi-size window fusion technology, A recursive window fusion algorithm is presented. After simulation, LIBSVM is transplanted to DM8127 for pedestrian discrimination, and the algorithm for extracting HOG features is transplanted. The whole intelligent monitoring system is implemented and tested. The experimental results show that adding the pedestrian with only head-shoulder information as the training positive sample can effectively solve the problem that the lower body of the pedestrian is occluded, and adjust the extraction parameters of the HOG feature. The proposed recursive window fusion algorithm can effectively transplant the algorithm combining multiple windows with hog and SVM into DM8127 to detect more than 90% of pedestrians.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】