Research On Visual Slam Based On Semi-direct Method

Humans’ exploration to the artificial intelligence has promoted the development of unmanned driving,unmanned aerial vehicles,robots,and has also brought new challenges to the field of computer vision.Simultaneous localization and mapping(SLAM)technology is the basis for the development of unmanned driving and other fields,and the camera has the advantages of low cost and rich image content.Using the camera as the sensor SLAM technology(visual SLAM,v SLAM),has become a research hotspot.The v SLAM algorithm based on the direct method is fast when building maps in large scenes and can handle weak texture scenes,but it is difficult to achieve map reuse and loop detection,and the cumulative error is large.The feature point method can achieve better when the texture information is rich,but the speed is slower.The semi-direct method combined with direct method and feature point method,with its unique advantages can construct a point cloud map through direct method matching,feature point method to loop detection and global optimization.The main research work includes:(1)A stereo SLAM method based on the combination of direct method and feature point method is designed.The stereo camera can use the left and right frames to solve the problem of uncertain scale and improve the speed of initialization.By combining the stereo direct method and the feature point method,using a stereo camera at the front end,by inputting the left and right frames of the image,the image is first tracked based on the direct method,that is,the stereo visual odometer DSO algorithm.The pose is checked in the active window continuously optimize.With the continuous increase of key frames,the marginalized frames of the stereo DSO are sent to the feature point method.This method can extracted the feature point and construct the bag-of-words model.Then,loop detection is performed.Finally the entire map is globally optimized to improve the robustness of the system.(2)The point cloud map constructed by the SLAM algorithm cannot be directly used for navigation and other applications.Based on the monocular semi-direct method LCSD-SLAM algorithm,the point cloud map established by the algorithm and combined with the ROS platform to transmit the pose and other information,the position can be obtained.The pose is calculated and transformed,and a 2D grid map is constructed,which can be used for navigation and so on in the future.