Monocular Visual SLAM Towards Dense Reconstruction

Simultaneous Localization and Mapping(SLAM)is one of the key technologies in the field of robotics,mainly focusing on localization and environment perception.Due to the unique advantages of visual sensors,visual SLAM has become a research hotspot in recent years.Feature-based camera tracking is the most widely used technology in monocular visual SLAM,and is capable of accurately tracking camera pose and inferring structure of the environment.However,features are heavily dependent on the environment,and are not performing well in textureless scenes.So,feature-based monocular visual SLAM always suffers from sparsity problem in reconstructed map,causing its sensing ability limited at the level of scene structure,which cannot be used in practical task,such as field scene detection and autonomous driving.Dense depth map estimation in non-structural area cannot be solved well only from the aspect of pure geometry theory.Convolutional neural network(CNN)has made a great progress on extracting high level feature and regression task of pixel level.In general,CNN-inferred depth is dense and globally accurate,which is complementary with structure data in process of camera tracking.Meanwhile,it has a better robustness because of characteristics of CNN architecture in low-texture scene.In this thesis,the monocular visual SLAM towards dense reconstruction has been addressed in term of depth fusion,which makes full use of feature-based camera tracking and CNN-inferred depth,and combines the traditional geometry methods with the concept of deep learning.The main research work includes:(1)We design and implement a complete Visual SLAM system towards dense reconstruction.An improved feature extraction strategy and wrong match culling algorithm are introduced to raise the accuracy of data correspondence.(2)In the dense depth estimation,an original CNN architecture with multi-scale module based on Resnet has been proposed,which enhances the performance of whole system under pure rotational camera motion and low-texture scene.(3)The scale ambiguity of monocular SLAM and uncertainty of depth map are well processed in the framework of depth fusion and reconstruction speeding up by a fast solver.