动作识别在交通指挥中的应用研究

发布时间：2018-01-23 00:08

本文关键词： 交通指挥关键帧姿态距离变换骨架提取模板匹配　出处：《河北大学》2017年硕士论文　论文类型：学位论文

【摘要】：伴随着科技力量和信息技术在社会发展中的不断投入,人民生活和社会经济都得到了飞速发展,而与之伴随的是越来越严重的城市交通拥堵现象。同时,无人驾驶汽车的问世以及智能驾驶辅助系统的不断升级发展,也对传统交通指挥形成了新的挑战。在人工智能和机器视觉领域中,动作识别技术正逐步趋于成熟,而应用于交通指挥中的动作识别仍然是一个新兴的、紧迫的研究热点。基于此背景本文主要针对交通指挥中的动作识别进行研究,主要研究内容如下:1.运用视频捕捉设备所录制的交通指挥动作,往往记载一定时间范围内的持续镜头内容,既包含了不同指挥动作的叠加,也有同一指挥动作在不同周期内的重复,这就需要采用运动特征变化率来进行运动分割。针对已经提取出运动目标区域的二值图像,首先需要获取其人体边缘,再根据其边缘变化引申出运动特征变化率,最后根据变化率的量化曲线界定动作分割。2.同一指挥动作中既包含具有决定性运动信息的关键帧姿态,也有对运动信息表达贡献较小的非关键帧姿态,这就需要采用表观特征轮廓向量来提取关键帧姿态。首先使用Candy算子提取目标区域的轮廓和质心,其次以质心到轮廓边缘的距离作为轮廓向量的元素,并且对轮廓变量进行推导优化,最后通过轮廓向量变化曲线定义关键帧姿态和非关键帧姿态,并完成关键帧姿态提取。3.针对交通指挥动作的特征提取,本文采用骨架参数作为选取特征,并对传统的基于距离变换的骨架提取算法提出优化和改进。传统的基于距离变换的骨架提取算法因离散域内最大内切圆的确定存在误差争议,以及骨架点的判断缺乏相邻点作为参考,这使得提取的骨架很难保证连通性及单像素性,本文通过定义种子骨架点并以生长模型搜索全部区域骨架点,不但保持了与原始图像一致的拓扑结构,并具有良好的连通性和单像素性,且对边缘扰动带来的噪声有很好的消除作用。4.本文针对常规模板匹配方法进行改进,以此重点突出手臂部位的信息表达。该方法以现行标准的交通警察指挥信号中提取的骨架模型制作样本模板,在利用豪斯多夫(Hausdorff)距离求取模板之间距离时,通过分析不同交通指挥动作骨架模型中身体各部位所贡献的运动信息差别,对不同区域设置相应权值,以达到提升交通指挥动作识别系统实时性、改善识别准确率的目的。
[Abstract]:With the continuous investment of science and technology and information technology in the development of society, people's life and social economy have been rapidly developed, and accompanied by more and more serious urban traffic congestion. At the same time. The advent of driverless vehicles and the continuous development of intelligent driving assistance systems have also posed a new challenge to traditional traffic command in the field of artificial intelligence and machine vision. The technology of motion recognition is becoming more and more mature, but the motion recognition used in traffic command is still a new one. Based on this background, this paper mainly focuses on the action recognition in traffic command, the main research contents are as follows: 1. Using video capture equipment recorded by the traffic command action. The continuous lens content in a certain time range is often recorded, which not only contains the superposition of different command actions, but also has the repetition of the same command action in different cycles. In view of the binary image which has been extracted from the moving target region, it is necessary to obtain the human body edge first. Then according to the edge changes, the rate of motion feature change is derived. Finally, according to the quantization curve of the change rate, the motion segmentation is defined. 2. The key frame posture with decisive motion information is included in the same command action. There are also non-key frame postures that contribute little to the expression of motion information, which requires the use of apparent feature contour vector to extract the key frame attitude. Firstly, Candy operator is used to extract the contour and centroid of the target region. Secondly, the distance from the center of mass to the edge of the contour is taken as the element of the contour vector, and the contour variables are derived and optimized. Finally, the key frame attitude and the non-key frame attitude are defined by the contour vector variation curve. And the key frame attitude extraction. 3. For the traffic command action feature extraction, this paper uses skeleton parameters as the selection feature. The traditional skeleton extraction algorithm based on distance transform is optimized and improved. The traditional skeleton extraction algorithm based on distance transform is controversial because of the determination of the maximum tangent circle in discrete domain. And the skeleton points lack of adjacent points as a reference, which makes it difficult to ensure connectivity and single pixel of the extracted skeleton. In this paper, the seed skeleton points are defined and the whole region skeleton points are searched by the growth model. It not only maintains the topological structure consistent with the original image, but also has good connectivity and single pixel. And the noise caused by edge disturbance is well eliminated. 4. This paper improves the conventional template matching method. This method uses the skeleton model extracted from the current standard traffic police command signal to make the sample template. Using Hausdorff-Hausdorff distance to calculate the distance between templates, the difference of movement information of different parts of the body in different traffic command action skeleton models was analyzed. In order to improve the real-time performance of the traffic command action recognition system and improve the recognition accuracy, the corresponding weights are set up for different regions.
【学位授予单位】：河北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U495;TP391.41

【参考文献】