Semi-direct Monocular Visual SLAM Method In Complex Environments

Visual simultaneous localization and mapping(SLAM)is a key technology in the research of intelligent robots.Traditional monocular visual SLAM algorithms are mostly built on feature-based methods or direct methods.These two methods have their own advantages and disadvantages.This paper proposes a semi-direct monocular visual SLAM algorithm,which combines the advantages of feature-based methods and direct methods.The main contributions of this paper are as follows:1)First,based on the framework of ORB-SLAM,this paper combines the advantages of feature-based and direct methods.The proposed method uses a direct method to track the camera pose quickly in the front end,which avoids the time consumption caused by feature extraction and matching.When a keyframe appears,the oriented fast and rotated brief(ORB)feature of the current image is immediately extracted and the feature matching is used to track the local map.Then the camera pose is optimized by minimizing the reprojection error.The ORB featurea of keyframes will also be used for creation of map points and optimization.The drift in long-term navigation of direct methods remains a problem.This proposed method makes use of keyframes for loop closure detection to eliminate accumulated errors.In addition,the ORB features of the current image will also be extracted during relocation and the camera pose will be estimated by using the feature-based method.Experiments have proved that the proposed method obtains better real-time performance while maintaining high accuracy.2)Secondly,traditional visual SLAM methods are designed for static scenes and have poor accuracy in dynamic environments.To deal with dynamic environments,this paper makes use of the data association between adjacent images to adopt a motion detection module that is robust to illumination change.The accuracy and robustness of localization and mapping is improved by removing moving points.The module establishes foreground probability of pixels through the temporal and spatial characteristics of pixels to find out the pixels belonging to the moving objects.This paper also uses the foreground probability to limit the sampling space of the model.The experimental results prove the effectiveness of the method.3)In loop closure detection,the proposed method creates the visual phrase set by considering the stable neighboring spatial relationship between two visual words and employs an improved pyramid term frequency-inverse document frequency(TF-IDF)scoring match scheme.Finally,visual words and visual phrases jointly determine the similarity between images.Experimental results show that the loop closure detection algorithm proposed reduces the perceptual aliasing and improves the recall of loop closure detection.