低空航拍下基于隐式姿态模型的平躺人体检测方法研究

项目名称： 低空航拍下基于隐式姿态模型的平躺人体检测方法研究

项目编号： No.61202143

项目类型： 青年科学基金项目

立项/批准年度： 2013

项目学科： 计算机科学学科

项目作者： 苏松志

作者单位： 厦门大学

项目金额： 25万元

中文摘要： 平躺人体检测是计算机视觉研究中的难点和热点，具有很好的理论意义，可广泛应用在灾害救助和家庭服务机器人等领域中。但目前低空航拍下平躺人体检测的研究处于起步阶段，存在视角变化大、姿态多样、背景复杂等尚未解决的问题。针对这些问题，本课题提出了低空航拍下基于隐式姿态模型的平躺人体检测方法，思路如下：首先,根据透视变换的几何含义，提出一种新的采样策略解决视角变化大的问题；其次，提出了隐式姿态模型来解决姿态多样化的问题，即构建人体的局部姿态检测器，基于回归森林对局部姿态的空间分布进行建模，通过局部姿态的投票来确立人体的位置；最后，提出了基于高斯过程回归的在线领域自适应策略，对局部姿态检测器的输出值进行动态更新，从而解决复杂背景的问题。该项目的研究拓宽了航拍图像的应用领域，有利于模式识别和计算机视觉等学科中相关理论和技术的发展，具有较大的理论意义和很好的应用前景。

中文关键词： 平躺人体；目标检测；计算机视觉；航拍图像；

英文摘要： Lying pose human detection is an active research field of computer vision in recent years. It has a good theoretical significance and many applications such as victim detection, home service robot. But the research on lying pose human detection in low-altitude aerial images is on its infancy, existing many unsolved problems due to the large variation of view-point, human pose, and image background. We focus on solving these problems, and propose a research on lying pose human detection in aerial images based on implicit pose model. Firstly, according to the geometry interpretation of perspective transformation, we propose an image sampling strategy to overcome the multi-view problem; Secondly, We propose an implicit pose model to deal with the large variation of pose, which determining the location of human by fusing the votes from local pose detector. The spatial distributions of all the local pose detectors are trained by regression forest. Finally, an online domain adaptation method based on Gaussian Processing Regression （GPR）, which used to update the scores of local pose detectors, is proposed to deal with the cluttered background. This research project is expected to deal with those mentioned problems; it can extend the application of aerial images, promote the development of relative theories and technol

英文关键词： lying pose human body；object detection；computer vision；aerial images；