Depth perception based on monochromatic shape encode-decode structured light method

Highlights

• The designed monochromatic pattern is robust against occlusion and sharp deformation, suitable for dynamic scene, and less sensitive for color interference, with symbol density spectrum (SDS) rule, legitimately increasing number of feature points inside element which further enhancing resolution.

• The proposed adaptive-symmetry-template (ASD) can robustly detect feature points.

• Based on transfer learning theory, using DenseNet to identify element label.

• Graph topological structure and look-up-table (LUT) help reducing decoding time.

Abstract

This paper presents a depth perception system based on one-shot coding structured light method. The designed monochromatic pattern is robust against occlusion and sharp deformation, suitable for dynamic scene, and less sensitive for color interference, with symbol density spectrum (SDS) rule, legitimately increasing number of feature points inside element which further enhancing resolution. For the decoding stage, firstly, the proposed adaptive-symmetry-template (ASD) can robustly detect feature points, secondly, in virtue of transfer learning, using DenseNet to identify element label, thirdly, graph topological structure and look-up-table (LUT) help reducing decoding time. The experiments show that the encoding method increases resolution fairly a lot and for several complex objects, the decoding method can achieve high accuracy and keep robustness.

Introduction

3D reconstruction and depth perception are continuous popular in the computer vision domain, among various ways of acquiring three-dimensional information, structured light owns advantages of non-contact, flexibility, easy control and high contrast [1,2]. Basic structured light system is composed of a projector and a camera, the projector projects single or multiple patterns carrying with coding information onto target while the whole depth perception process, i.e. decoding, is accomplished based on information captured by camera. According to distinct coding strategies of projected pattern, structure light methods can be divided into spatial and temporal categories [3].

Temporal coding combines the camera with dynamic projected pattern, even it can obtain pixel-by-pixel disparity, it also faces challenges such as motion artifacts in dynamic scenes and poor instantaneity [4,5,6]. For these problems existing in temporal coding, Sam et al [7] innovatively used deep neural network to effectively solve. The spatial coding especially for one-shot coding strategy is possible for real-time 3D depth perception while the coded pattern plays a significant and intuitionistic role in improving resolution, accuracy and precision [8,9,10], the decoding algorithm decides precision and speed of final depth perception [11,12,13]. Compared with temporal coding, spatial coding is relatively weak in resolution and accuracy.

To conquer deficiencies and combine existing preponderances, this paper adopts one-shot pattern with particularly designed approaches to enhance resolution and accuracy. Based on the De Bruijn sequence to generate array, the intersections of each embedded element are considered to be primary feature points, the elements are especially designed with secondary feature points using SDS rule. To robustly and accurately decode pattern, transfer learning, adaptive-symmetry-template, graph topological structure and LUT are bonded.

The rest paper is organized as follows. Section 2 introduces the related work. Section 3 introduces specific implements of encoding and decoding methods. Section 4 shows the experimental results, analysis and comparison with related literature. Conclusion is drawn in last section.