Local RGB-D Feature Detection,Feature Description And Its Application

Local feature extraction is usually the first stage of the computer vision or im-age processing tasks,such as wide baseline matching,image stitching and image classification,so the property of the local feature will influence the system's perfor-mance.With the fast development of the RGB-D sensor,the consuming-level RGB-D sensor becomes popular gradually.Compared with the traditional camera,RGB-D sen-sor can capture the depth information directly which adds a greatness to the RGB-D local feature,so designing a better RGB-D local feature is an advanced project.Local feature extraction contains three steps:feature detection,feature description and feature matching,so researching the three comprehensive is a skillful study.In this thesis,based on the existing RGB-D descriptor LOIND(Local Ordinal In-tensity and Normal Descriptor)[1],we introduced a new RGB-D detector and an optimal implementation of the RGB-D descriptor.In addition,we collect a benchmark to evalu-ate the local feature completely and accurately.The main contributions are as follows:1.We mount the RGB-D sensor on a high precision robotic arm,complete the eye-in-hand calibration and collect the RGB-D benchmark in different scenes.The benchmark contains 3 main classes and 15 subclasses which can be used to eval--uate the local feature completely and accurately.2.A RGB-D detector is introduced which uses RGB information and depth infor-mation based on the same method of Harris detector[2]respectively and then com-bines them to solve the problem that Harris fails to detect the interesting points under a pool light condition.In addition,the RGB-D detector can help improve the performance of the RGB-D descriptor.3.The methods used to solve the scale estimation and orientation estimation are introduced which help the RGB-D feature adapt to the scale transformation and rotation transformation,and then we optimize the algorithm details and code to improve the performance of the RGB-D local feature.