Simultaneous Localization And Mapping Based On Improved Optical Flow

Simultaneous Localization and Mapping technology is one of the research hotspots in the field of mobile robots,drones,and automatic driving.SLAM technology can help robots,drones,and unmanned vehicles to independently determine their position,to perceive information from the surrounding environment,so as to achieve autonomous control and navigation and other functions.As one of the most promising SLAM methods,visual SLAM has attracted more and more attention.This article focuses on the visual SLAM method,uses Kinect1.0 and a MYNT camera as a visual sensor to study visual SLAM.Based on the traditional optical flow method,this paper proposes an improved optical flow method based on GPU acceleration.Based on this improved optical flow method,we propose an RGBD SLAM system and improve the VINS Mono.The main content of this article is as follows:(1)Study the relevant background knowledge of SLAM,including the development history of SLAM,and optimize the SLAM method based on graph after the SLAM sparsity is proposed.Introduce the relatively outstanding open source SLAM system and the successful commercial application of SLAM,and put forward the research focus of this paper.(2)Study the background knowledge of visual SLAM algorithm,including the camera model,how Microsoft Kinect 1.0 perceives depth information;introduce the rigid body motion in three-dimensional space and how to use Lie group Lie algebra to optimize the position and pose;conducted research of the optical flow method of calculation methods and principles.(3)Based on the traditional optical flow method,an improved optical flow method with GPU acceleration is proposed.Based on the acceleration of GPU operation,the operating speed of the improved optical flow method is improved.By judging the distance of the feature points after the conduct optical flow twice,the erroneous feature points are eliminated,and the correctness of the tracking of the feature points is improved.Compared with the ordinary optical flow method,the average operating time of the optical flow method is improved from 0.0636 s to 0.0026 s,the tracking accuracy rate is increased from 70% to 95%,and it has strong robustness in various complex scenarios such as light changes,lack of features,blurred images,and dark images.The SLAM system lays a good foundation for its operational efficiency and robustness.(4)Based on the improved optical flow method,an improved optical flow method RGBD SLAM system is proposed,which includes front-end visual odometer and backend local optimization and loop close detection.In this system,the front end extracts the ORB feature points,uses the improved optical flow method to track the feature points,calculates the camera pose,and uses the tracking rate of the feature points to filter the key frames;the back end selects some key frames and performs Local Bundle Adjustment optimizes the three-dimensional position of the feature point and the pose of the camera at the same time.The loop close detection section detects whether loop close occurs and optimizes the global pose of the camera after the loop close occurs to obtain an optimized global camera position.After the test on the TUM dataset and the actual office scene test,the accuracy of feature point tracking is above 90%,compared with the description of the sub-matching method,the average image processing time per frame is 0.0287 s,and we can processe 30 images per second.With this SLAM system,the number of key frames is 5% of the initial frames,which greatly reduces the number of key frames.The accuracy of loopback detection is 100%,and the positioning accuracy is slightly improved compared to RGBD SLAM.(5)Based on the improved optical flow method,the Visual Inertial Fusion open source VINS Mono is improved.Using a MYNT camera as a visual sensor,the camera and the IMU are calibrated,and the data is collected in the field for a positioning test.By testing in EuRoC datasets and real-world scenarios,VINS Mono optimized by the improved optical flow method has improved the positioning accuracy by about 20% over the original one.