Non - Real - Time Three - Dimensional Reconstruction Of Rotating Objects Based On Kinect

In recent years, visual technology and virtual reality technology developed constantly, multimedia technology has been used in-depth. Currently many mature technologies have come out of laboratory, being widely used in the areas such as industrial manufacturing, medical services, education and entertainment. Due to many applications’ pursuing of real sense, needs for three dimensional reconstruction of real-world objects are increasing rapidly. While traditional manual modeling is relatively inefficient and slow although it can guarantee the quality and preserve the inner structure of objects. The old methods can hardly meet the increasing needs. Now there has been some technologies invented for three dimensional reconstruction based on laser scanners and binocular vision. The former technology is expensive both in price and data storage. And the later one is very complex in computation and data organization. A cheap, precise and easy method is needed to accomplish3-D reconstruction.Using Kinect device, I have done some research on this problem and got a method for non-real-time3-D reconstruction based on color image data and depth data.This thesis is mainly about three dimensional reconstruction based on Kinect data and explores several problems.1. To get the target to be reconstructed, I used the background subtraction method using a static background image based on color image data to ensure the quality of the result. And proved that GMM method cannot be applied for target extraction with experiments. Either, it is impossible to use the depth data for target extraction.2. In the problem of feature detection in target regions, I got a statistical relationship between the algorithm threshold and the number of features detected through the process on a series of sample data, and got a proper solution for deciding thresholds by analyzing the result.3. To balance the need for quality and efficiency of the feature matching stage, I found a solution to sample several frames of the original data to reduce calculation. 4. During the reconstruction stage, a transform method between depth data domain and pixel units was applied under the condition that the precision requirement could compromise. Reconstruction was made possible by unifying two data types. Finally, the model was joint together using the Bursa-wolf method and the result of feature matching.Concluding the experiment result, it’s not hard to see that3-D reconstruction based on Kinect device is much easier compared to traditional methods. While at the same time, the precision of the model is somewhat sacrificed. It is still impossible for real-time process because of the complex processing steps and calculation, but sometimes a good choice for low-precision usage.