基于引导采样的Kinect深度图修补算法

Kinect depth map retrieval based on guided-sampling

针对Kinect采集到带有大量的结构性缺失的深度图,提出了一种基于引导采样的深度图空洞噪声修补算法。算法首先将深度图所对应的彩色图片转换为灰度图,然后用K-means算法将彩色图转换而来的灰度图进行聚类处理,将生成的聚类图作为引导图。联合引导图对深度图空洞噪声边缘深度值采样,采集多个深度值并计算深度均值,最后使用深度均值来作为空洞的深度估计值。通过与基于蒙特卡罗不确定度评价的深度图修补算法(MC-UE)相比较,由于有引导图的矫正作用,边缘细节更加清晰准确。对于处理较小面积的空洞噪声,处理结果相较于MC-UE算法,均方误差降低4%左右。对于处理较大面积的空洞噪声,均方误差较MC-UE算法降低了9.65%~14.32%。实验证明引导采样算法在处理较大面积空洞噪声时相较于MC-UE算法有更低的均方误差。

Focusing on the depth map with large amount of structural defects collected by Kinect, this paper proposed an algorithm for patching noises in the hole of the depth map based on the guiding sample. Based on this algorithm, it converted the color picture corresponding to the depth map to grey-scale map at first, which was then treated using clustering processing with the K-means algorithm, making the clustering picture produced as the guiding map. Combining the guiding map, it sampled depth values of the noise edge in the hole of the depth map, the mean value of which was then calculated and was used finally as the estimated depth value of the holes. Through comparison with the depth map patching algorithm based on the Monte Carlo uncertainty evaluation（MC-UE）, the edge details are clearer and more accurate due to the corrective effect of the guiding map. For the noises in the hole with fairly small area, the treatment of which leads to a decrease of the mean squared error by nearly 4 % compared to that by MC-UE algorithm. While for the noises in the hole with large area, the treatment of which results in a decrease of the mean squared error by 9.65 %-14.32 % compared to that by MC-UE algorithm. Ths experiments demonstrate that the guiding sample algorithm contributes to lower mean squared error in the treatment of the noises in the hole with large area than that by MC-UE algorithm.