复杂背景下视频人脸检测方法研究

发布时间：2018-01-12 07:40

本文关键词：复杂背景下视频人脸检测方法研究　出处：《海南大学》2016年硕士论文　论文类型：学位论文

【摘要】：人脸检测与跟踪在视频监控、人脸识别、表情识别、性别识别、年龄识别、医疗诊断、人机交互等领域有广泛的应用前景。虽然目前存在大量的人脸检测算法,但是人脸图像会受到各种因素的影响,其中包括：光照,视角,姿态,年龄老化,表情,发型,脸部饰物等。人脸的非刚性和环境的复杂性,使得即使是同一人在不同条件下也存在巨大的检测差异,所以想用一种算法来进行所有情况的人脸检测是不可能的。本文对运动目标检测算法进行了改进,并且融合肤色特征、几何特征、模板匹配三种人脸检测算法进行分层过滤,实现了静态图像和动态图像的人脸检测。其中复杂背景主要包括背景光照变化,树木扰动,噪声等情况。动态图像包括网上现有的视频序列和以固定摄像机拍摄的实时视频序列。具体工作如下：(1)对视频序列进行了预处理,分别使用参考白算法、均值滤波法进行光照补偿、平滑去噪。(2)针对视频中的冗余背景信息,本文通过比较光流法,帧间差法,背景减除法的优势、局限性和适用场合,采用了改进的三帧差法和混合高斯模型建模的背景减除法相结合的算法,克服了单一检测方法的局限性。并将本文的方法与其他方法进行实验对比。(3)分析比较色彩模型和肤色模型后,用改进的YCbCr色彩空间构建阈值分割模型,再将分割出的肤色区域进行形态学滤波,连通区域标定。用几何知识滤除连通区域中非人脸区域,再对候选区域进行人脸模板匹配,从而实现人脸的精确定位。(4)详细介绍了Mean Shift算法、Cam Shift算法和KLT跟踪算法的原理,最后使用改进的P-KLT算法进行人脸跟踪,并将本文的方法与Cam Shift算法进行对比实验,证明了本文改进的方法的有效性。
[Abstract]:Face detection and tracking in video surveillance, face recognition, expression recognition, gender recognition, age recognition, medical diagnosis, human-computer interaction and other areas have a wide application prospect. Although there are a large number of face detection algorithm, but the face image will be affected by various factors, including: light, angle, attitude, age aging, facial expression, hairstyle, facial ornaments. The face of non rigid and the complexity of the environment, so that even if the same people also have great differences in the detection of different conditions, so I want to use a face detection algorithm for all situations is not possible. This paper presents an improved algorithm for moving target detection. The fusion and color characteristics, geometrical feature, template matching three face detection algorithm for layered filtering, realized face detection static and dynamic images. The complex background mainly includes the background light Change, trees and so on. The noise disturbance, the dynamic image including the existing online video sequence and shot by a fixed camera video sequence. The specific work is as follows: (1) the video sequence is processed, using white reference algorithm, mean filtering method for illumination compensation, smooth denoising for (2). Redundant background information in the video, by comparing the optical flow method, frame difference method, background subtraction, advantages, limitations and applications, using the improved three frame difference method and modeling of mixed Gauss model combining the background subtraction algorithm, overcomes the limitation of single detection method. And this paper the experiment compared with other methods. (3) analysis and comparison of color model and color model, construct the threshold segmentation model using YCbCr improved color space, the skin color segmentation by morphological filtering and connected region Calibration. With the knowledge of geometry filter connected non face regions in the region, the candidate region of the face template matching, so as to achieve precise positioning of the face. (4) introduces the Mean Shift algorithm, Shift algorithm and the principle of Cam KLT tracking algorithm, finally using the improved P-KLT algorithm for face tracking, and the Cam method and Shift algorithm are compared, proved the effectiveness of this improved method.

【学位授予单位】：海南大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】