在线获取场景的三维结构信息以及对场景中动态目标进行三维测量是计算机视觉技术的研究重点，也是虚拟现实、智能视频监控、机器人以及无人驾驶的核心技术，然而，如何提高其计算效率仍然存在问题。本文针对在线获取场景三维结构和动态目标测量技术开展深入的研究，主要贡献如下：

（1）针对视觉SLAM中大场景的稠密重建问题，本文以RGB-D为相机输入设备，以RTAB-Map为系统框架，开发了一套在线稠密3D重建系统。实验结果表明，本文中所采用的基于RGB-D相机和RTAB-Map系统框架，通过双边滤波改善Kinect原始深度图质量，提高了RTAB-Map系统中相机跟踪的精度，并且针对SLAM中的闭环检测提出了一套高效的内存管理策略，严格控制闭环检测以及图优化中关键帧的数量，极大提高了计算效率，同时降低了内存消耗，可以应用在移动设备的大场景三维重建中。

（2）提出了一种基于单目视频的单人体目标高度的实时在线测量方法。该方法首先利用高斯混合模型（Gaussian Mixture Model）获得背景图像，然后通过当前帧图像与背景图像的差分获得运动人体区域的前景图像，之后对前景图像进行形态学处理以消除图像碎片区域，得到较完整的前景图像后，根据人体形态特征在前景图像中提取头部特征点、脚部区域以及脚部特征点，根据获取的图像特征点建立约束方程求取动态人体目标的高度。实验表明，该方法能够准确的计算序列图像中的运动人体的高度，并且该算法不需要复杂的计算就可以实现实时测量，此外，算法的稳定性高，可以不需要精确的前景和双脚检测结果，对于分辨率较低的视频序列仍然能够得到较为准确的高度测量值。

（3）提出了一种基于单目视频的多人体目标高度的实时在线测量方法。该方法首先利用Vibe（Visual Background Extractor）算法分割出运动人体目标前景模型，再通过团块检测和扩展卡尔曼滤波(Extended Kalman Filter)算法确定视频中每一个运动人体的矩形区域，然后在每一个目标的矩形区域提取人体的头部特征点和脚部特征点，最后通过建立约束方程求解每一个动态人体目标的高度。实验结果表明，该方法采用的背景建模算法可以只通过一帧图像快速的实现前景和背景的分割，即使出现光照突变或者背景变化的情况，也能快速准确的提取出前景模型，并且在确定单个人体目标区域时，采用基于扩展卡尔曼滤波的目标跟踪算法确定人体区域的搜索范围，可以提高算法的计算效率，还能防止由于分割出的人体前景区域不连通而导致特征点检测错误。

Online 3D reconstruction and dynamic object measurement are important tasks in the field of computer vision, which has many applications such as in virtual reality, intelligent video surveillance, robotics and unmanned driving. However, efficiency of  the computations is still a problem. This thesis studies online implementations of 3D reconstruction and dynamic object measurement. The main contributions are:

1. An online large scale dense visual SLAM (Visual Simultaneous Localization and Mapping) system based on RGB-D camera is developed. The system is modified based on RTAB-Map system. Experimental results show that the accuracy of camera tracking is better than RTAB-Map system because of application of bilateral filter on original Kinect depth map. The memory management strategy of the system can strictly control the number of key frames, which are suitable for loop closure detection and pose graph optimization. Besides, this system is efficient in computations and reduces memory consumptions. This method can be applied to mobile devices for large scale 3D reconstruction.

2. A method is proposed to measure the height of a moving human from a video sequence in real time and online. Firstly, this method extracts the background model of the video by Gaussian Mixture Model(GMM). Secondly, it obtain the foreground image of the human body by image difference between current image and the background model. Then, eliminated the image area of debris in foreground image are by morphological processing. After that, a key point on the head image area, foot image area and key points on the foot image area are extracted from the foreground image. Finally, the height of the human body is computed by several geometry constraint equations. Experimental results show that the proposed method can compute accurately the height of the moving human from a video sequence, and is real-time which has lower computational cost. Moreover, it is robust to inaccurate segmentations and low resolutions.

3. A method is proposed to measure the height of multiple moving human from a video sequence in real time and online. Firstly, this method extracts the foreground model of human bodies by Vibe(Visual Background Extractor) algorithm. Secondly, it obtains the rectangle area of each human body by blobs detection and object tracking with extend Kalman filter. Then, head feature point and foot feature point from each rectangle area of human body are extracted. Finally, the height of each human body in the video is computed by several geometry constraint equations. Experimental results show that this method can subtract the background model by only one frame, and it is robust to sudden change of light or background. Besides, it is computational efficiency to locate each rectangle position of human bodies and robust to obtain feature points from fragment of foreground, which is due to the advantage of object tracking algorithm for reducing the search range.