基于ARM的人员智能引导系统的设计

发布时间：2018-05-20 16:20

本文选题：智能引导系统 + ARM　；参考：《东华大学》2010年硕士论文

【摘要】： 根据视频统计现场人数并进行自动引导,是智能视频监控新的应用趋势。近年来,以上海世博会为代表的大型会展的兴起为智能引导系统(IGS)的发展提供了契机。人们开始思考如何在嵌入式系统中运用图像处理技术,实现人数的自动识别,进而完成智能引导,以提高参观效率,保障游客安全。智能引导系统具有很高的商业价值和发展潜力。本文综合运用嵌入式、图像处理等技术,从大型展厅的实际需求出发,设计了基于ARM的人员智能引导系统,采集现场图像并实现人员自动计数,系统分为硬件和识别算法两部分。系统的硬件选取EmbedSky公司的ARM9开发板TQ2440作为开发平台,采用CMOS图像传感器OV9650采集现场图像。在Linux环境下编写摄像头驱动程序,实现图像采集,并将采集后的图像生成BMP位图文件。对采集的位图图像通过梯形低通滤波器,调整阀值以去除大部分高斯噪声。引入信噪比(SNR)作为判断参数,判定样本图像是否需要去噪,以减少运算量;对图像中因相对运动而产生的运动模糊,使用投影恢复法予以消除;采用背景差法提取前景目标,提出背景自适应算法更新现场背景,抑制现场光照变化对目标检测的影响。使用Canny边缘检测算法寻找边缘,提取轮廓。依据图像形态学理论,反复进行腐蚀、膨胀运算,填充封闭的空白区域,消除人员遮挡和粘连带来的误判,去除与人员无关的微小点,形成较为完整的封闭轮廓;基于Hu矩不变量,构造了基于物体轮廓线的轮廓矩不变量,它具备平移、旋转和尺度不变性,独立于物体本身的灰度。通过分析模板图像的轮廓矩不变量,设定允许误差的阈值,与现场提取的样本图像的轮廓矩进行匹配,判断人员数量,结合展厅容量以及报警设定值实现自动引导。识别算法在Visual C++下,使用OpenCV编写,调试通过后编写Makefile文件,移植到Linux系统。实验结果表明,在光线充足且背景相对简单的场景中识别准确率较高,当光线减弱或者人员与背景纹理相近时,识别准确率出现下降。
[Abstract]:It is a new application trend of intelligent video surveillance to count the number of people in the field and conduct automatic guidance according to the video. In recent years, the rise of large-scale exhibition, represented by Shanghai World Expo, has provided an opportunity for the development of Intelligent guidance system (IGS). People begin to think about how to use image processing technology in embedded system to realize the automatic recognition of the number of people, and then complete the intelligent guidance, in order to improve the visit efficiency and ensure the safety of tourists. Intelligent guidance system has high commercial value and development potential. According to the actual demand of the large exhibition hall, this paper designs an intelligent personnel guidance system based on ARM, which can collect the scene images and realize the automatic counting of the personnel. The system is divided into two parts: hardware and recognition algorithm. The hardware of the system selects the ARM9 development board TQ2440 of EmbedSky company as the development platform, and adopts the CMOS image sensor OV9650 to collect the field image. In the environment of Linux, the camera driver is written to realize the image acquisition, and the captured image is generated into the BMP bitmap file. Through trapezoidal low-pass filter, the threshold is adjusted to remove most of the Gao Si noise. SNR (SNR) is introduced as the judging parameter to determine whether the sample image needs denoising in order to reduce the amount of computation; to eliminate the motion blur caused by relative motion in the image, the projection restoration method is used to remove it; and the background difference method is used to extract the foreground target. A background adaptive algorithm is proposed to update the scene background to suppress the effect of the field illumination change on the target detection. Canny edge detection algorithm is used to find the edge and extract the contour. According to the theory of image morphology, repeated corrosion, expansion operation, filling the closed blank area, eliminating the personnel occlusion and adhesion caused by misjudgment, removing the tiny points unrelated to the personnel, forming a relatively complete closed contour; Based on Hu moment invariant, the contour moment invariant based on object contour is constructed. It has the invariance of translation, rotation and scale, and is independent of the gray level of the object itself. By analyzing the invariant of the contour moment of the template image, setting the threshold of the allowable error, matching the contour moment of the sample image taken from the field, judging the number of the personnel, combining the capacity of the exhibition hall and the alarm setting value to realize the automatic guidance. The recognition algorithm is written with OpenCV under Visual C, and then the Makefile file is compiled after debugging, and then transplanted to Linux system. The experimental results show that the recognition accuracy is higher in the scene with sufficient light and relatively simple background. When the light is weakened or the person is close to the background texture, the recognition accuracy decreases.
【学位授予单位】：东华大学
【学位级别】：硕士
【学位授予年份】：2010
【分类号】：TP273.5

【引证文献】