Accurate feature point detection method exploiting the line structure of the projection pattern for 3D reconstruction.

Abstract

The 3D imaging methods using a grid pattern can satisfy real-time applications since they are fast and accurate in decoding and capable of producing a dense 3D map. However, like the other spatial coding methods, it is difficult to achieve high accuracy as is the case for time multiplexing due to the effects of the inhomogeneity of the scene. To overcome those challenges, this paper proposes a convolutional-neural-network-based method of feature point detection by exploiting the line structure of the grid pattern projected. First, two specific data sets are designed to train the model to individually extract the vertical and horizontal stripes in the image of a deformed pattern. Then the predicted results of trained models with images from the test set are fused in a unique skeleton image for the purpose of detecting feature points. Our experimental results show that the proposed method can achieve higher location accuracy in feature point detection compared with previous ones.