基于神经网络的RGB-D图像物理量非接触式精确测量

本课题将以RGB-D图像为输入，利用神经网络技术分析目标物体结构和成分，非接触式的精确测量目标物体的多种物理量。现有方法无法在某些场景中有效的测量，例如：无法直立者或传染病人的身高体重；运动、失重状态下物体质量；强放射性物体表面材质与摩擦因数；搬运机器人快速获知物体质量和密度等。 本课题拟研究深度图像对物体拓扑结构的表达，设计一种三维重建方法，分析深度图像缺失特征及彩色图像有效特征，实现表面材质识别和三维模型准确分割，计算摩擦因数；研究物体结构和组成成分，计算密度信息；研究深度图像增强方法，估计体积和质量。本课题拟结合计算机图形学、机器视觉、计算机辅助设计等交叉学科，为准确快速的非接触式估计未知物体质量、密度、体积、摩擦因数等多种物理量提供理论模型和技术方案。仅使用商品级相机可减少设备支出便于应用推广。只采集一张图像可减少时间成本，满足疫情防控、虚拟现实、机器人、自动驾驶等领域的应用需求。

This project application proposes to take RGB-D images as input, use neural network technology to analyze the structure and composition of the target object, to eventually accurately measure and estimate multiple physical quantities of the target object in a non-contact fashion. Conventional methods cannot effectively measure under some circumstances such as, body height and weight of patients who are contagious or not able to hold upright; the mass of objects in motion or under weightlessness; surface material and friction factor of highly radioactive objects; handling robots to quickly sense the mass and density of objects, etc. This topic propose to study the effect of depth images on the topological structure expression of objects to design a 3D reconstruction method; analyzing the missing features of depth images and the effective features of RGB images, implementing surface material identification and accurate segmentation of three-dimensional models, calculating friction factors; studying object structure and composition, calculating density and other information; studying depth image enhancement methods, estimating the volume and mass. The subject is intended to provide theoretical support and technical solutions for fast, accurate and non-contact measure or estimation of various physical quantities such as mass, density, volume, friction factor, etc, in combination with computer graphics, machine vision, and computer-aided design. The fact of only using commodity RGB-D cameras can reduce the equipment expenditures to facilitate application. Also that collecting only one image help reduce time costs and meet application needs in scenarios including epidemic prevention and control, virtual reality, robotics, and autonomous driving.