A depth image acquisition platform based on Kinect V2

Abstract

In this manuscript, an acquisition platform for depth images based on Kinect V2 is designed, which can acquire depth images of the target model at any attitude angle (including view angle 0~80°, azimuth angle 0~360° and spin angle 0~360°). In addition, this manuscript implements a depth image recognition algorithm based on an integrated local surface patch (LSP). The algorithm first calculates feature points in regions with large shape variations, and then defines a LSP at each feature point, which is characterized by its surface type, the patch centroid, and the 2D histogram. Next, the potential corresponding patch pairs are found by matching two sets of LSPs, and the candidate models are obtained by the filtered potential corresponding patch pairs. Finally, the candidate models are validated by the iterative closest point (ICP) algorithm. Experiments are designed to validate the performance of the algorithm using multiple depth images with different attitude angles and occlusion ranges of eight military target models acquired by the platform. The results show that this depth image acquisition platform can provide rich data support for the design and verification of depth image recognition algorithms in the future.