移动相机下基于三维背景估计的运动目标检测

发布时间：2018-01-19 21:41

本文关键词： 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　出处：《仪器仪表学报》2017年10期 　论文类型：期刊论文

【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。
[Abstract]:Moving target detection in indoor environment is a hot topic in the field of computer vision. The dynamic background caused by mobile camera is the difficulty of moving target detection. In this paper, a moving target detection algorithm based on synchronous location and map creation ORB-SLAM 3D background estimation is proposed. Firstly, the mobile camera is used to traverse the whole indoor environment, and the ORB-SLAM technology is used to establish the 3D background feature point cloud model of the current global environment. Then the local 3D feature point cloud is built based on the local video, and the current local 3D feature point cloud and the environment 3D background feature point cloud are embedded according to the location information, which is based on the environment background information. The 3D mean shift (3DMS) algorithm is used to extract the foreground feature points of the local 3D feature point clouds. Using the deep convolution neural network, the candidate areas of the extracted foreground feature points are confirmed by the moving targets. The results are verified by the actual experiments of many indoor scenes. The results show that this method has a high accuracy and recall rate of moving target detection. The proposed algorithm makes full use of 3D background information and uses deep convolution neural network to confirm. The accuracy and robustness of the detection are improved effectively.
【作者单位】：中国科学院深圳先进技术研究院;中国科学院大学深圳先进技术学院;
【基金】：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
【分类号】：TP391.41
【正文快照】： 0引言运动目标检测是计算机视觉中的一项重要技术[1],主要应用领域包括视频监控、视频编码、目标识别与跟踪、行为分析等[2-4]。此外,该技术经常被用于其他计算机视觉相关工作的预处理当中,如视频抠图[5]等。传统的移动目标检测技术假设相机是固定静止不动的,近年来研究人员提

【相似文献】