基于多特征的行人快速检测方法研究

发布时间：2018-10-29 09:30

【摘要】：当前计算机视觉领域研究的热点之一就是人脸识别和行人检测,这一技术已经被广泛的应用在很多领域,比如智能电话、智能交通、无人驾驶等。由于算法的精度和速度等原因,很难应用到实时系统中。传统方法中为了提高精度需要计算图像特征金字塔,而构建特征金字塔需要花费大量时间。本文正是从这一问题出发,通过估算多尺度分类器的方法,减少构建特征金字塔时间从而提高了行人检测的速度。本文主要的研究内容如下:(1)本文提出了一种新特征BPG。BPG特征实质是HOG的一种变形,它既保留了原有的梯度和方向信息,也具有局部区域信息。BPG特征把梯度方向分成8个方向,在可变大小区域内的不同梯度方向上累加梯度值,然后与均值作比较,进行二值编码,最后生成十进制数。实验结果显示,新特征在行人检测方面有更强的识别能力。(2)通过实验对比和特征间的相关性挑选出四个特征作为特征池,四个特征分别是BPG特征、LBP特征、梯度特征值和下梯度方向特征。根据特征本身的特点分析了特征间的互补性。用这四个特征融合对单一样本的检测正确率大约为97%。(3)分类器设计主要使用的是Adaboost算法。通过级联形式,使每级强分类器有不同数量的弱分类器和阈值,使分类器的识别能力逐级加强。这样可以将容易识别出的负样本在第一级或是前几级分类器就排除掉,难以识别的由后面识别能力强的分类器识别,既提高了分类器的精度又减少了检测窗口的数量。(4)提出了一种估计分类器的方法,通过估计临近放缩层的分类器以代替图像的缩放过程,大幅度减少计算特征金子塔所耗费的时间,实验表明估计分类器算法虽然会使检测的精度下降1-3%,但检测的速度却提升了 2倍多。
[Abstract]:Face recognition and pedestrian detection are one of the hot topics in the field of computer vision. This technology has been widely used in many fields, such as smart phones, intelligent transportation, driverless and so on. Because of the accuracy and speed of the algorithm, it is difficult to be applied to real-time system. In order to improve the accuracy of traditional methods, we need to calculate the image feature pyramid, but it takes a lot of time to construct the feature pyramid. From this point of view, the method of estimating multi-scale classifiers is used to reduce the time of constructing feature pyramids and improve the speed of pedestrian detection. The main contents of this paper are as follows: (1) in this paper, we propose a new feature, BPG.BPG feature, which is essentially a kind of deformation of HOG, which not only preserves the original gradient and direction information. The BPG feature divides the gradient direction into eight directions and accumulates the gradient value in the different gradient directions in the variable size region. Then compared with the mean value the binary coding is carried out and finally the decimal number is generated. The experimental results show that the new features have stronger recognition ability in pedestrian detection. (2) four features are selected as feature pool through experimental comparison and correlation among the features. The four features are BPG features and LBP features. Gradient eigenvalues and lower gradient directional features. The complementarities between features are analyzed according to the characteristics of the features themselves. The accuracy of detecting a single sample with these four features fusion is about 97. (3) the classifier is designed mainly using Adaboost algorithm. By cascading, each strong classifier has a different number of weak classifiers and threshold values, and the recognition ability of the classifier is enhanced step by step. In this way, the negative samples that are easy to identify can be excluded at the first or first stages of the classifier, and those that are difficult to recognize are identified by the classifier with strong recognition ability behind them. It not only improves the accuracy of classifiers but also reduces the number of detection windows. (4) A method of estimating classifiers is proposed to replace the zooming process of images by estimating the classifiers near the scaling layer. The experimental results show that the estimation classifier algorithm can reduce the accuracy of detection by 1-3 and increase the speed of detection by more than 2 times.
【学位授予单位】：内蒙古大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】