The Implementation Of The Depth Sensing System

With the development of computer vision,3D data is widely adopted in human-computer interaction,3D reconstruction,biomedical science,machine vision,etc.In depth sensing field,it is difficult to simultaneously achieve depth maps with high accuracy and large resolution in real time.In this paper,we construct a depth sensing platform based on infrared projection,and design an accelerating algorithm on GPU platform according to the depth sensing method proposed by Li Q et al^[1].The algorithm achieves real-time performance in experiment.We firstly introduce the method proposed by Li Q et al^[1],which consists of two main steps:speckle matching and phase interpolation.The method is capable of achieving high accuracy with the cost of computation expensive.In order to solve the time-consuming problem,we propose a pipeline of GPU based algorithm to accelerate the implementation of the two steps.In the speckle matching process,one needs to calculate the normalized corre lation coefficients?NCC?of each pixel block in the captured image with all the pixel blocks in the pattern.The calculating operation reaches10¹³.To accelerating the calculation,we firstly decrease the needless computation by splitting the formula of NCC,and secondly reduce the memory access latency by optimizing data storage strategy.In the phase interpolation process,the phases are obtained by performing a 2D filtering on the captured speckle image with a Gabor filter.However,the 2D filtering is also computation expensive.After singular value decomposition of the Gabor kernel,we observe that only one singular value is large and the others are approximate to zero.With the aid of SVD,we transform the 2D filtering into two 1D convolutions,which decreases the computation dramatically.Experiments show that the proposed technique is able to achieve excellent performance on accuracy,resolution and real time implementation.By adopting infrared illumination,the system is robust to the in-door ambient light.The technique provides a new solution to human-computer interaction,machine vision,etc.