三维空间中基于图结构的人体姿态估计算法研究

Human pose estimation has attracted wide attention due to its important application value and theoretical significance. However, confounding factors such as the diversity and complexity of body posture, occlusion and background clutter, make human pose estimation in the real scene is far away from matured. Therefore, the algorithm of human pose estimation based on computer vision remains a challenging problem. To end this, the depth information and body pose invariance method has been proposed in complex scene to improve the accuracy and reliability of human pose estimation. The main problems this study concerns on could concluded as following: (1) To research on 3D invariance feature extraction for human pose description and tracking; (2) To design a method for occlusion detection and part reconstruct based on 3D Space, and use this method to reduce the error in human pose estimation; (3) To explore human pose estimation method based on graph matching in complex scene for the realization of the diversity pose estimation. The research on this topic discuss and explore the challenge issues in object detection and recognition, it can promote the practical application of computer vision, therefore has important theoretical and practical values.

人体姿态估计以其重要的应用价值和理论研究意义引起了广泛关注。然而人体姿态本身的多样性和复杂性，以及场景中的遮挡和背景干扰等因素使得真实场景中的人体姿态估计算法尚未成熟，仍然是一个充满挑战的研究领域。因此，本课题将深度信息引入到人体姿态估计中，以实现复杂场景下准确可靠的多样化人体姿态估计，具体拟开展以下三方面工作：（1）研究三维空间的不变性特征，用于姿态变化中的人体部位描述及跟踪；（2）设计基于三维空间的部位遮挡识别及其恢复方法，用于解决由遮挡引起的姿态估计错误问题；（3）探索复杂场景中基于图结构的人体姿态估计方法，用于实现多样化人体姿态估计问题。本课题的研究有助于推动计算机视觉研究的实际应用，同时对目标检测与识别中的难点问题展开探索，既有广泛的应用价值，同时又具有十分重要的科学意义。

人体姿态估计以及动作分析一直是计算机视觉研究的热点和难点，有着广泛的应用价值。而复杂场景中遮挡、背景干扰和人体动作引起的姿态形变等因素会给姿态估计和动作分析带来极大挑战。本项目针对这些问题，围绕人体这一核心目标，以人体姿态估计为切入点，融合人体骨架模型，场景信息以及视频中的运动信息等，提高姿态估计和人体动作分析算法的性能.具体地，本课题在以下几个方面取得了一定的成果：（1）研究对比了主要的人体姿态估计算法，发表了人体姿态估计的综述文章；（2）研究了基于人体姿态的入口尾随进入检测算法；（3）研究了基于场景上下文的人体行为分析算法；（4）研究了基于多平面投影特征的人体行为分析算法；（5）研究了一种基于稀疏表示和半监督回归的人数估计算法。本研究对视频监控、动作分析等领域有较高的学术与应用价值。