项目名称: 基于多光谱成像的人体头肩检测系统研究
项目编号: No.61301184
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 无线电电子学、电信技术
项目作者: 侯亚丽
作者单位: 北京交通大学
项目金额: 24万元
中文摘要: 人体检测在HOV车道(即大容量车辆行车道)上的乘客检测、人机交互及未来视频监控等应用中都有着非常重要的作用。但是,目前多数检测系统面临背景中形似人体物品的干扰。多光谱图像可以获得比传统RGB图像更多的场景信息,将有助于降低系统误检率。头肩检测是人体检测中一个重要部分,基于目前研究现状,本项目将从三个方面展开基于多光谱图像的头肩检测系统研究。首先,本项目将依据皮肤的独特反射特性提出一种对多光谱头肩检测系统进行有效波段选择的算法,简化成像系统复杂度;第二,研究多光谱图像融合算法,综合利用单像素反射特性与空间纹理特征提高多光谱皮肤及头发区域检测的可靠性,为头肩检测提供更有力的依据;第三,研究基于多光谱图像的头肩检测特征,并分别探讨不同光照条件下的特征提取方法,利用多光谱信息提高检测性能。最终,将头肩特征检测与皮肤及头发区域结果相结合,建立一个多光谱头肩检测系统。
中文关键词: 多光谱成像;活体检测;卷积神经网络;热红外成像;疲劳驾驶
英文摘要: Human detection is an important component in HOV(High Occupancy Vehicle) occupant detection, human-robot interaction, video surveillance, etc. However,in most current systems,objects with appearances like human beings may easily cause false detections in real scenarios.Compared with the conventional grayscale/RGB iamges,more information about the scene could be obtained in multispectral images, which could be used to reduce the false alarms. Head-shoulder is an important part of human detection.Based on an extensive study of the previous work,three aspects about head-shoulder detection in a multispectral imaging system will be further studied in this project.First, effective bands for the head-shoulder detection system will be selected based on the reflection curves of human skin. A new band selection algorithm will be developed. By using the selected bands, the imaging system can be simplified and the feature dimensions can be reduced. Second, a method to use both reflection properties of each pixel and texture properties of local regions will be developed for skin detection in multispectral images. Methods about image fusion will be studied to extract texture features from multispectral images.The skin and hair region detection results will provide more reliable cues for head-shoulder detection. Third, featu
英文关键词: multispectral imaging systems;liveness detection;convolutional neural networks;thermal images;fatigue driving