2D转3D视频系统中运动检测算法的研究及实现

发布时间：2018-08-13 08:27

【摘要】：随着电视技术的发展,3D视频显示技术渐渐走入人们的视野,人们可以在屏幕上观看到立体的影像,并在军事,医疗,教育等领域得到快速发展。一种获得3D视频的方法是通过2D转3D视频技术直接将2D视频转为3D视频,这种方法制作成本低周期短,而且有丰富的2D视频资源可供转换,因此受到广泛研究。在2D转3D视频系统中最为关键的是深度图提取,本文主要研究基于运动检测的深度图提取的方法,这是一种将运动目标当作前景的深度赋值方式,这种假设的方式符合观影的习惯特点且具有很高的效率,运动目标检测结果就成为这种方法的核心所在。然而大多数的运动检测结果受环境的影响很大,如果场景中存在动态背景就会产生大量的误检点,运动物体运动速度过快就会产生鬼影,运动物体遮挡光线则会将阴影区域也检测为运动目标,针对以上可能出现的问题,本文提出一种准确有效的运动检测算法。针对鬼影和场景中存在动态背景的问题,提出一种基于时间和空间信息的背景模型,这种背景模型能够有效去除大部分的误检点;减少漏检点的方法则是在前后景分割的过程中结合一种“或”类型的三帧差法,并将分割结果和显著性检测结果相与,可以进一步去除误检点;最后基于HSV颜色空间去除阴影区域。在Matlab和Opencv平台对算法进行软件层验证,实验结果表明算法可以得到正确的运动检测结果,量化指标f-measure可以达到0.80以上。
[Abstract]:With the development of television technology, 3D video display technology has gradually come into people's view, people can watch stereoscopic images on the screen, and get rapid development in military, medical, education and other fields. One way to get 3D video is to convert 2D video to 3D video directly through 2D to 3D video technology. This method is widely studied because of its low cost, short period and abundant 2D video resources. In 2D to 3D video system, the key is depth map extraction. This paper mainly studies the method of depth map extraction based on motion detection, which is a kind of depth assignment method which regards moving target as foreground. This hypothetical method is consistent with the habit of viewing and has high efficiency. The result of moving target detection is the core of this method. However, most of the results of motion detection are greatly affected by the environment. If there is a dynamic background in the scene, there will be a large number of false detection points, and if the moving object is moving too fast, it will produce a ghost image. The shaded light of moving object will also detect the shadow area as moving target. In view of the above possible problems, this paper proposes an accurate and effective motion detection algorithm. A background model based on temporal and spatial information is proposed to solve the problem of dynamic background in ghosts and scenes. This kind of background model can effectively remove most of the false detection points. The method of reducing missed detection points is to combine a "or" type of three frame difference method in the process of scene segmentation, and to combine the segmentation result with the significant detection result, which can further remove the false detection points. Finally, shadow area is removed based on HSV color space. The software layer of the algorithm is verified on the platform of Matlab and Opencv. The experimental results show that the algorithm can get correct motion detection results, and the quantization index f-measure can reach more than 0.80.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN948.6;TP391.41

【参考文献】