Research On Hybrid Sparse Monocular Visual Odometry

Simultaneous Localization and Mapping(SLAM)and Visual Odometry(VO)are the problems of constructing a map of the surrounding environment and determining the current location of the sensor.They are widely used in unmanned vehicles and ARs.Augmented Reality),VR(Virtual Reality),and autonomous robots,with the popularity of artificial intelligence,VO/vSLAM plays an increasingly important role in people's lives.Most monocular VO/vSLAM algorithms focus on the feature-based methods or the direct methods,while the semi-direct(hybrid)method is often ignored as an equally important method.Based on the existing semi-direct method framework,this paper redesigns the whole system and completes a novel hybrid sparse monocular visual odometry method.The on-line radiation calibration of the camera based on the sliding window is integrated into the system,which calibrates the camera's response function,vignetting,and exposure time online,making the algorithm robust to strong camera vignetting and sudden changes in exposure time;With a large number of white walls or other weak texture conditions,the number of corner points in the image is often sparse.Due to the lack of information,this will bring great difficulties to the general algorithm.In this paper,the corners and edges are combined.It has been added to the system to make the system run stably in weak texture environment;the VO/vSLAM based on semi-direct framework tends to have high efficiency,but it is lacking in accuracy and robustness.The direct framework has been modified to obtain a hybrid sparse framework,which runs through the whole system.On the one hand,the system has better efficiency than the general feature-based method and the direct method.On the other hand,on the basis of the semi-direct method,Improving the accuracy and robustness of the system;the system adopts a keyframe extraction strategy combining relaxation and caustics,which is as strict as possible under normal circumstances.The extract keyframes,thus reducing the redundant information to improve the accuracy of the algorithm,as quickly extract keyframes in extreme environments,which ensures the robustness of the algorithm.The proposed system is compared with ORB-SLAM,DSO and SVO2.0 on the EuRoC MAV,ICL-NUIM,TUM mono VO,TUM-RGBD datasets,and the results show that the proposed Hybrid Sparse monocular visual Odometry(HSO)achieves better or comparable performances than state-of-the-art monocular algorithm.