Research On Real-time Image Processing Algorithm In Mobile System

In recent years,with the development of science and technology and the improvement of people's living standard,intelligent mobile systems such as smart cars,unmanned aerial vehicles,sweeping robots come into people's lives,mobile systems that based on image enhancement become more and more popular.This paper focuses on mobile systems,especially computer vision algorithms used for cars.In this paper,some commonly used image processing algorithms and SLAM algorithms are studied,and the main research results are as follows:(1)to analyze the characteristics of LSD-SLAM and ORB-SLAM,and to compare with the advantages and disadvantages of theirs,in this paper,three different data sets are tested with LSD-SLAM and ORB-SLAM respectively,and compare the results.It is found that LSD-SLAM is stable in the case of slow motion and small change between frames,and it can produce semi-dense point cloud.However,when the motion is fast or the change between frames is large,the LSD-SLAM is prone to tracking failure.The ORB-SLAM in all cases have excellent performance,and there is a good loop closure detection and relocation function,and ORB-SLAM usually runs faster.However,if the camera moves too slow,it will produce more key frames,so that the running speed will decrease.In addition,the 3D point clouds obtained by ORB-SLAM are more sparse,they are difficult to see the specific outline of the object,and they are more suitable for the track record and macroscopic analysis.(2)combined with optical flow tracking and ORB-SLAM technology to achieve a single camera 360-degree aerial view of the panorama.This paper first analyzes the shortcomings of the traditional method which uses four cameras to achieve the aerial view of the panorama,it is of high cost and the installation is complex.So this paper proposes a solution that using a single camera in the car and generate the aerial view of the panorama after moving a distance.Firstly,the original image is undistorted and projected to a top view,and then the optical flow is tracked.Since the corresponding point satisfies the homograph with 3 degrees of freedom,the homograph matrix can be calculated by solving the linear problem,and the aerial view can be generated.However,there are still fewer feature points on the ground,and the points above the ground do not satisfy the transformation relationship.This paper presents a solution based on ORB-SLAM algorithm to realize the aerial view.Through we can obtain the position of the camera,and then get the homograph matrix between two frames in the top view,then the aerial view of the panorama can be generated.(3)combined with visual geometry algorithm and ORB-SLAM technology to achieve pseudo-3D panorama.In this paper,we first analyze the shortcomings of the aerial of the panorama,and propose the algorithm of generating the pseudo-3D panorama through the images collected by the four cameras.Then we use the "bowl" model and the three planar 3D projection model to obtain the pseudo-3D panorama.Because the centers of the four cameras are not at the same position,this algorithm just can realize a pseudo-3D panorama.Due to the selection of projection models and the parameter settings,there will be stitching errors and other issues.Experiments show that the "bowl" model is better in a spacious road scene,and in a narrower and crowded road scene,the three planar projection model is more appropriate.On this basis,this paper combines the ORB-SLAM technology to obtain the cloud points near the vehicle,and then uses RANSAC to fit planes which can be used to estimate the distance of the surroundings from the car,and then adjust the parameters of the three planar projection model,finally,we can get the most suitable projection model to generate the adaptive pseudo-3D panorama.