SIA OpenIR  > 智能检测与装备研究室
Alternative TitlePassenger Car Rate Intelligent Video Detection Algorithm
Thesis Advisor马钺
Keyword车载率 边缘识别 边缘识别 Rhtrht 发色特征 纹理svmsvm
Call NumberTP183/C45/2013
Degree Discipline计算机应用技术
Degree Name硕士
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract本文针对长途客车运行中存在的超载等三超三私等问题,通过建立车载视频监控系统,完成对长途客车的实时监控和管理,而在车载视频监控系统中,乘客的检测算法设计是其最重要的软件设计环节。在充分考虑长途客车的实际情况后,在垂直视角下完成乘客识别,将实际问题转化为俯视图像中乘客识别的问题。 在视觉上以及实际采集到的图像可知,俯视图像中乘客头部轮廓较为清晰明显,所以本文综合考虑算法的复杂度和实时性要求后,选取随机Hough变换检测算法,并对传统的随机Hough算法做出许多改进,小单元化了边缘点,候选圆心聚集分组,分布化最优拟合轮廓提取等,建立在本课题研究应用基础上的改进随机Hough变换检测算法(URHT)。实验表明URHT算法可以准确的检测出乘客的头部区域。检测精度和实时性都有很好的表现。 我们在继续研究乘客俯视图下的头部信息时发现,发色的颜色深度较大,并且具有明显区别于其他区域的纹理特性,这提供给我们大量新的头部特征信息。我们引入了机器学习的思想,将支持向量机的算法原理应用到头部发色和纹理区域特征检测的问题当中,将原来的检测问题转换为头部目标的最优化分类识别问题。我们选取多项式核函数,训练发色纹理分类器,经过多次训练之后,发色纹理分类器可以表现出很好的分类效果。在URHT检测算法和发色纹理分类器级联检测的前提下,我们将视频图像序列的图像帧间信息加以利用,用混合Gauss模型对背景进行建模,获取前景区域和背景区域。将感兴趣区域的范围大大缩小,进一步提高了检测的准确性和实时性。 本文实验在光照条件较好条件下,完成实际乘客登车视频图像的检测。实验结果表明,本文算法的时间复杂度能够满足检测实时性的要求,在检测准确率上可以达到90%以上,能够满足实际的需求。
Other AbstractIn this paper, aiming at the existence of long-distance bus safety management problems, on board vehicle monitoring platform is established to complete the calculation of rate of car passengers. In the building of vehicle monitoring platform, the key problem is how to solve the identification of the passengers when getting on the vehicle. During the process of getting on, the camera is fitted on the top of the coach door, thus the figure acquired is top view, i.e. the so-called vertical viewing angle. The passenger detecting problem in this case comes down to the identification of the passenger’s heads. In the top view, the area of the passenger’s head has many obvious feature and the detection just starts from these features. The first thing is to obtain the contour of head visually and significantly. In the vertical perspective, the passenger’s head is circle-like which inspired us to adopt the circle or circle-like detection algorithm to complete the task of head detection. Considering the complexity and accuracy of the algorithm, the RHT transform detection algorithm is selected to accomplish the goal. However, traditional RHT detection algorithm cannot complete the task in this paper. Based on this algorithm, a lot of experiments were done which indicated that the URHT algorithm can detect the passenger's head accurately to give the exact size and the location information, but the background circle-like target may cause false detection. Continuous study on the top view of passenger’s head shows that hair color is depth and texture characteristics significantly different from other regions. This provides a large number of new features about the head, so we have introduced the idea of machine learning. The algorithm which supports vector product is applied to the problem of head hair color and texture region feature detection. As a result, the detection problem is converted to a head target optimization problem of classification and recognition. The hair color texture classification is trained with C-SVC (C-support vector product) and the polynomial kernel function. After several training the classification showed good classification results. With the cascaded detection of the URHT detection algorithm and hair color and texture classification, the inter-frame information specific to the video image and Gauss modeling method is used to modeling the background to get the prospect area and the background area. Thus the region of interest is greatly reduced, and the detection accuracy is improved as a result. The experiments were done in good lighting conditions with the actual long-distance bus boarding detection video. Experimental results show that the time complexity of the algorithm is able to meet the requirements of real-time detection. The detection accuracy can reach more than 90% which meets the actual demand.
Contribution Rank1
Document Type学位论文
Recommended Citation
GB/T 7714
陈昊. 客流车载率智能视频检测算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2013.
Files in This Item:
File Name/Size DocType Version Access License
客流车载率智能视频检测算法研究.pdf(2028KB) 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[陈昊]'s Articles
Baidu academic
Similar articles in Baidu academic
[陈昊]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[陈昊]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.