视频中多视点人脸检测技术的研究

发布时间：2018-05-03 05:09

本文选题：人脸检测 + 多视点　；参考：《深圳大学》2017年硕士论文

【摘要】：社会治安一直是民生生活中一大难题。而对于人口密集,流动性大的城市,治安问题更为严重。出于电子技术的发展,各种社会治安场合已经安装了摄像头进行监控,籍此缓解治安压力。事发后,视频资料是可提供破案线索的重要信息。即便如此,视频资料的人脸信息收集仍然需要人力资源。通过电脑检测则可以很好的自动化这一过程,但对于检测技术来说,一些特殊情景仍然难以处理:比如对于非完整目标的检测;再比如从设施安放来看,摄像头的位置可能被人为移动,造成旋转、部分遮挡等等也会对检测的效果造成干扰。以往的检测方式在直立的脸检测较好,而对发生旋转、部分遮挡的目标检测效果则不好。出于此类原因,本文训练了一个在旋转、遮挡方面有较强适应性的检测器。在视频检测中利用视频的时空间一致性信息改进检测效果,并训练了一个卷积神经网络用来估计不同视点下的人脸姿态。具体贡献如下:首先,本文基于卷积神经网络训练了一个人脸检测器,并对其进行了优化改进。优化网络模型设计,减少需要学习的参数量。利用样本扩充技术、改善样本分布提高学习效果。引入非完整的人脸作为弱正样本,以解决遮挡问题。对于弱正样本与正样本得分相同而使得依赖得分和重叠率的非极大值抑制算法失效的问题,本文提出利用不均衡样本的策略去增强正样本的分值。其次,视频中存在运动模糊、遮挡、形态变化、光照变化等,仅仅使用图片检测手段不能取得很好的效果。本文利用视频的特性如时空间一致性、运动场等信息,平滑检测效果,与未利用视频特性的检测器在效果上进行对比。再者,对于多视点问题,不同视点下所观察到人类的姿态有着不同的俯仰、朝向、旋转。传统的姿态评估方法通常是通过回归的方法得到人脸的一些关键点的位置,如眼睛、鼻子、嘴巴,然后用一个3D的人脸模型的不同姿态下的2D投影使得模型上的关键点与2D投影中的关键点尽可能的匹配,以此来得到人脸的姿态。本文训练一个卷积神经网络,直接得到人脸的姿态,避免了制作3D模型、与3D模型进行匹配的过程。
[Abstract]:Social security has always been a major problem in people's livelihood. For densely populated, mobile cities, the security problem is even more serious. Due to the development of electronic technology, cameras have been installed in various social security situations to relieve the pressure of public order. After the incident, video data can provide clues to solve the case important information. Even so, the face information collection of video data still needs human resources. Computer testing is a good way to automate this process, but for detection techniques, some particular scenarios are still difficult to handle: for example, detection of non-complete targets, or, from the point of view of facility placement, The position of the camera may be moved artificially, causing rotation, partial occlusion and so on. In the past, the detection method was better in upright face, but bad in target detection with rotation and partial occlusion. For this reason, we have trained a detector with strong adaptability in rotation and occlusion. In video detection, the temporal and spatial consistency information of video is used to improve the detection effect, and a convolution neural network is trained to estimate the face pose at different viewpoints. The specific contributions are as follows: firstly, a human face detector is trained based on convolutional neural network and optimized. Optimization of network model design to reduce the number of parameters to learn. Using the technique of sample expansion to improve the distribution of samples and improve the learning effect. In order to solve the occlusion problem, a nonholonomic face is introduced as a weak positive sample. For the problem that the weak positive sample has the same score as the positive sample and makes the non-maximum suppression algorithm dependent on score and overlap rate invalid, this paper proposes a strategy of using unbalanced samples to enhance the score of positive samples. Secondly, there are motion blur, occlusion, morphological changes, illumination changes and so on in the video. Only using image detection methods can not achieve good results. In this paper, the features of video, such as time-space consistency, sports field and other information, are used to smooth the detection effect, and compared with the detector without video characteristics. Furthermore, for the multi-view problem, the human attitude observed at different viewpoints has different pitching, orientation and rotation. Traditional attitude assessment methods usually use regression method to get the position of some key points of the face, such as eyes, nose, mouth, Then 2D projection of a 3D face model is used to make the key points of the model match the key points of 2D projection as far as possible to get the pose of human face. In this paper, we train a convolutional neural network to get face pose directly and avoid the process of making 3D model and matching 3D model.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】