摘要
在视频中把人体对象从复杂背景中自动分割出来是一个非常有挑战性的问题。对此提出了一个新的人体对象自动分割方法,可以在相对复杂的背景下分割出人体对象。首先利用Kinect对视频中的人体骨架信息进行跟踪,然后把骨架信息映射到高质量摄像头拍摄的视频中,最后利用跟踪到的骨架信息对高质量图片中的人体对象进行分割。为了得到这个映射关系,利用SURF特征进行匹配,根据匹配的体征点,采用最小二乘法估计映射参数。最后,利用映射后的人体骨架信息,对图片进行标记,然后采用Lazy-Snapping算法对图片进行人体对象分割。实验结果表明,该方法在640×480分辨率的视频中对人体对象的自动分割可以达到20帧/秒,而且精度与文中提到的其他算法相当。
Automatically extracting human objects from complex background in video images is a very challenging problem. A new approach for human segmentation which can extract human objects from complex background has been proposed in this paper. Firstly, we obtain the tracked human skeleton data in video images with Kinect. Secondly, map them to another image space generated by a HD camera which is capable of providing qualified images. Finally, split human subjects from the high-quality pictures based on the skeleton information. In order to get this mapping relationship we use SURF to do feature match, and estimate the mapping parameters using least squares method according to the signs point of the match. Then, mark the picture according to the human skeleton after mapping. At last the Lazy-Snapping method is adopted to segment the human in the video with the skeleton mark. Experiments conclude that the method can get a segmenting frame rate up to 20 FPS in 640 × 480 resolutions video and approximate accuracy as other algorithms mentioned in the paper.
出处
《电子测量技术》
2013年第4期48-51,共4页
Electronic Measurement Technology