郭会文,吴新宇,苏士娟,傅睿卿.移动相机下基于三维背景估计的运动目标检测[J].仪器仪表学报,2017,38(10):2573-2580
移动相机下基于三维背景估计的运动目标检测
3D Background estimation for moving object detection using a single moving camera
  
DOI:
中文关键词:  室内运动目标检测  三维背景估计  三维均值漂移  深度卷积神经网络
英文关键词:3D background estimation  moving object detection  3D Mean Shift (3DMS)  deep convolutional neural networks
基金项目:国家自然科学基金面上项目(61473277) 、机器人与智能制造国家地方联合工程实验室(2015581)、深圳市智能机器人与智能制造工程实验室(20141722)、深圳市技术攻关项目(JSGG20150930154605341)资助
作者单位
郭会文 1.中国科学院深圳先进技术研究院深圳518055;2.中国科学院大学深圳先进技术学院深圳518055 
吴新宇 1.中国科学院深圳先进技术研究院深圳518055;2.中国科学院大学深圳先进技术学院深圳518055 
苏士娟 中国科学院深圳先进技术研究院深圳518055 
傅睿卿 中国科学院深圳先进技术研究院深圳518055 
AuthorInstitution
Guo Huiwen 1. Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen 518055, China; 2. Shenzhen College of Advanced Technology, University of Chinese Academy of Sciences, Shenzhen 518055, China 
Wu Xinyu 1. Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen 518055, China; 2. Shenzhen College of Advanced Technology, University of Chinese Academy of Sciences, Shenzhen 518055, China 
Su Shijuan Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen 518055, China 
Fu Ruiqing Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen 518055, China 
摘要点击次数: 1255
全文下载次数: 1427
中文摘要:
      室内环境中的运动目标检测是计算机视觉领域的研究热点,而移动相机造成的动态背景是运动目标检测的难点。本文提出一种基于同步定位与地图创建(ORB SLAM)三维背景估计的运动目标检测算法,首先使用移动相机遍历整个室内环境,采用ORB SLAM技术建立当前全局环境的三维背景特征点云模型;然后基于局部视频建立局部三维特征点云,根据定位信息将当前局部三维特征点云与环境三维背景特征点云进行嵌入,基于环境背景信息,采用三维均值漂移(3DMS)算法,对局部三维特征点云进行前景特征点提取;运用深度卷积神经网络,对提取的前景特征点所在候选区域进行运动目标确认。通过多个室内场景的实际实验进行验证,结果表明本文方法具有较高的运动目标检测准确率和召回率,提出的运动目标检测算法充分利用了三维背景信息,采用深度卷积神经网络进行确认,有效地改善了检测的准确性和鲁棒性。
英文摘要:
      Moving object detection in indoor environment is a research hotspot in the field of computer vision. The dynamic background caused by moving the camera is a challenge in moving target detection. In this paper, a moving object detection algorithm based on ORB SLAM (oriented FAST and rotated BRIEF Simultaneous Localization and Mapping) is proposed. Firstly, the entire indoor environment is traversed using a moving camera, ORB SLAM is used to establish the 3D feature cloud model of the global background. Then, based on the environmental information, local 3D feature point cloud is built. By embedding the local 3D feature points into the global 3D background feature cloud, 3D Mean shift is applied to extracting the foreground points from the local 3D feature points. Finally, deep convolution neural network is utilized to confirm the moving target of the candidate region where the extracted foreground feature points are located. The experimental results on multi indoor scenes show that the proposed method has high detection accuracy and recall rate. The proposed moving object detection algorithm makes full use of the background information, and the depth convolution neural network is used to confirm candidate regions, which effectively improves the detection accuracy and robustness.
查看全文  查看/发表评论  下载PDF阅读器