Combining Feature And Direct Methods For Robust Vision SLAM

In recent years,with the continuous development of computer vision technology,simultane-ous localization and mapping based on vision has become a hotspot in the field of robotics.In visual SLAM,the visual odometer used to estimate the position and posture of continuous frames is the most important one.In visual odometer,the inter-frame pose estimation mainly uses two methods:feature method and direct method.Feature method,which has long been considered as the mainstream method of visual odometer,makes full use of the robustness of feature points to light,scale,noise and rotation.And After years of development,it is more mature in matching and pose calculation.However,because it only extracts the obvious feature points in the environ-ment,it has poor adaptability to low texture environment,and only can constructe a sparse point feature map.Direct method is more robust to low-texture images because it does not need a lot of information with obvious features in the image,and because it saves the time of calculating feature points and descriptors,it has less computation and faster speed.However,because direct method is based on the assumption of photometric consistency,it is sensitive to illumination,and most of them need to calibrate the camera's photometric parameters.Direct method can still be used in the case of blurred image motion.but when the image has geometric noise,the performance degrades rapidly,and its robustness to large baseline motion is poor.Combining the advantages of both feature method and direct method,this paper proposes a view based on feature and direct method fusion.Sensory SLAM method can be used for real-time robot localization and mapping in indoor and outdoor environments.The main research results of this paper include:1.On-line photometric parameter estimation based on feature method.In order to improve the robustness of illumination,we extract the obvious feature points in the image,calibrate the exposure time of each frame on-line in real time,and estimate the response function and halo parameters of the camera on-line at the back end.2.Front-end tracking based on hybrid pyramid.A hybrid hierarchical inter-frame estimation al-gorithm is designed,which utilizes the advantages of the multi-pyramid strategy from coarse to fine in DSO to design a three-layer hybrid pyramid.First,extracting a small number of features and descriptors at the highest level for matching,and calculating the re-projection error.The coarse pose estimation between two frames is obtained to prevent excessive inter-frame motion and improve the robustness to fast motion.Then,the initial estimation of pose based on feature matching is used for iterative estimation of direct method,and the final accurate estimation of pose is obtained.3.Robust stereo visual odometry implementation.The direct matching method is used to achieve stereo scale recovery.In the back-end sliding window optimization,stereo scale esti-mation factor is added to improve the accuracy of scale estimation.4.Closed-loop detection and relocation based on feature method and direct method.A com-plete and unified SLAM system which combines direct method and feature method is de-signed and constructed,and a closed-loop detection and posture optimization and relocate system is added to the visual odometer.Experiments show that:The system has good posi-tioning accuracy and robustness,and can build a dense visual point cloud map.