Research And Implementation Of Stereoscopic Visual SLAM Based On Semi-direct Method

In recent years,emerging technologies such as intelligent mobile robot,driverless,augmented reality(AR)and virtual reality(VR)have attracted much attention at home and abroad.Driven by strong market and application demand as well as huge commercial potential,Simultaneous Localization and Mapping(SLAM)is constantly developing towards the direction of high real-time,high robustness and high precision,so as to provide high-quality location services and environment map for upper applications.The stereoscopic Visual SLAM(VSLAM)system has obvious advantages over the monocular VSLAM system,which not only avoids the scale ambiguity brought by monocular vision sensor,but also can build more vivid environment map.In this thesis,the stereoscopic VSLAM is researched and discussed,and a semi-direct method VSLAM is proposed to make full use of the advantages of stereoscopic visual.This method combines the advantages of direct method and feature-based method,which not only has the advantages of fast speed and low demand for scene texture,but also maintains the advantages of high accuracy and strong stability of feature method,and uses the Bag of Visual Word(Bo VW)model for loop-closed Detection and relocation,finally,a complete VSLAM system with a three-thread structure is realized,which is called SDF-SLAM(Semi-Direct Full SLAM).The main contributions of SDF-SLAM are:1.The system supports two mainstream stereoscopic vision sensors: stereo camera and RGB-D camera,which expands the system's scope of application and greatly improves the speed and success rate of the system initialization.In addition,a stereo ranging subsystem based on NCC(normalized cross correlation)measurement,especially for the stereo vision sensor,can calculate the feature depth from the stereo image quickly and accurately.2.Improved the ORB feature extraction algorithm by using the keypoint detection strategy based on the quad tree structure to improve the original ORB feature easy to gather on the image,so that it can adapt to the characteristics of VSLAM.Based on the improved ORB feature,a Visual Odometry(VO)based on semi-direct method was proposed.With the advantages of the feature method,the processing speed is maximized while maintaining accuracy.3.Based on the Bo VW model,K-means++ algorithm was used to train 42 different scenes and more than 200,000 pictures to obtain a huge ORB visual dictionary and use it for loop detection.Based on the ORB visual dictionary,the closed-loop detection module of SDF-SLAM system is realized,which effectively solves the loop-closed detection problem and eliminates the accumulated error to some extent.4.The SDF-SLAM system is evaluated and analyzed through experiments on the open-source dataset and the real-world environment dataset collected by individuals,and compared with many famous open-source VSLAM systems,which proves that the SDF-SLAM system has excellent realtime performance and practicability under the premise of maintaining good accuracy.