The Research Of Sample-based Learning Method For Human Depth Estimation In A Single Camera

The acquisition of image depth information is a basic problem and an intermediate link for computer vision. With the deepening of research, the technology of depth image acquisition has derived a large number of applications in real life, such as artificial intelligence Interactions, intelligent monitoring systems, medical aided diagnoses and the analyses of human movement, etc.The conventional manners to obtain the depth image can be broadly divided into two types:methods based on active depth sensors and methods based on computer stereo vision. They both have its drawbacks severally:the former one needs special equipment; the latter one needs a series of images or an image sequence. Aiming at problems existing in the traditional way to obtain depth images, this paper focuses on the researches of human depth estimation in a single camera from the perspective of Statistics.This paper has proposed a method to estimate human depth based on sample-learning in a single camera. The basic idea is:to establish a body depth database, to study human feature, to estimate and optimize human depth from the similar samples obtained by feature matching.The basic problems involved include how to establish human depth database, human contour edge detection,matching contour edges, selecting the most similar samples and human depth optimization. The main works and innovation points are:1. Aiming at the problem on detection of human contour edges, this paper discusses how to detect the human body contour edges from the gray images of the depth database and the color images of the single camera collection.2. Two kinds of human body contour matching algorithms are proposed. The first describes the similarity of two human body contour images, by calculating the distance between human body contour edge points to the center of gravity, extracting the outline about the edge of the multidimensional feature vector, and calculating the distance between the multidimensional vectors. The other one measure the similarity by distance transformations and edge distance. The former is simple, but the matching process is relatively rough. Low accuracy and mismatching are likely to occur. The latter, of which matching process is relatively complex, is more accurate and can be a good match to the similar samples.3. On the basis of the human body contour matching, this paper has proposed a sample-based learning method for human depth estimation in a single camera. On Xiamen University depth database, we perform an experiment to estimate human depth of color images obtained by a single camera. Finally, the simulated experimental results were found to validate the effectiveness of the proposed two methods.