SLAM Technology Research Based On Monocular Vision

Simultaneous Location and Mapping(SLAM)is one of important parts in AI research field and the key technology of self-driving.Also,it is the precondition to achieve self-localization and navigation of robots and has comprehensive application in both civilian and military.On the background of monocular SLAM,we have researched the front of visual SLAM,which is visual odometry and loop closure by using deep learning.Besides that,we put more effort to research the semi-direct monocular visual odometry algorithm and loop closure based on context learning and deep hashing and achieved some improvement,which can be summarized as below.We analyze the related technology of visual odometry and loop closure.Firstly,the feature-based visual odometry and the direct visual odometry are introduces in detail.Then,loop closure is analysed from two aspects which include the traditional methods based on visual vocabulary and the advanced methods based on deep learning and we give the merit and demerit at the same time.We study an algorithm about semi-direct monocular visual odometry based on fixed maps.Firstly,we introduce the SVO algorithm which first successfully fuse feature-based methods and direct methods together.Then,SVO algorithm is analyzed from three main parts which include the pipeline?motion estimation and reconstruction.After this,we give the disadvantages of this algorithm.For the initialization of SVO algorithm,we propose the fixed map to replace incremental map to accelerate the initialization and as for error accumulation of SVO,we propose to introduce relocalization into the pipeline to reduce the error drift.Finally,we test the proposed method on KITTI dataset and EuRoC dataset and get improvement in robustness and motion estimation precision.We research the loop closure by using feature learning and hash coding.Firstly,we introduce the motivation ? the frame of deep learning ? feature learning and fusion between the learned feature and hand-crafted feature to analyze the algorithm and point out its disadvantages.Then,for the two stages of feature learning and hash coding,we propose an end to end deep learning frame so as to reduce multiple quantization error causing by two-stage learning.To improve the speed of original algorithm,we cluster the position of the hand-crafted features thus the features belong to the same class can share one context,in this way,we can improve the speed of the algorithm remarkably.Finally,we test the proposed algorithm on New College dataset and TUM dataset and the results suggest that our method achieve a certain improvement in computing speed and precision recall.Specially,the computing speed is almost doubled.