With the rapid development of information technology and the proliferation of multimedia data, visual Simultaneous Localization and Mapping(SLAM) technology is gradually becoming a core technology for many applications, such as visual navigation, augmented reality, and street view modeling etc. Although comprehensive studies on visual SLAM systems have been undertaken, two key problems, the accuracy of mapping and the efficiency of map optimization, have been not resolved completely yet, which severely degrade the level of application of visual SLAM in an urban environment. Addressed these two problems, a systematical research has been conducted. The main contributions are: (1) A new map optimization method, which incorporates prior knowledge of an urban structure, is proposed. The proposed method first extracts the structured regions of a map, and then incorporates these structures by means of a new way of parameterization where the sparsity is preserved. Subsequently, a new sparse bundle adjustment method is present, which preserves the geometric invariance of these structured parts, and thus the accuracy of the built map is improved. Experimental results show that both the map and the camera motion are recovered with higher accuracy in the presence of prior information, compared with the methods incorporating no prior knowledge. The improvement is especially significant in the cases where only a few views are available or the baselines among the views are narrow. (2) A graph-partition-based bundle adjustment method is proposed, which exploits the sparsity of large scale maps of urban applications. By representing the locality of all the extracted local regions of the whole map with an undirected graph induced from the matched features among images. First, via the multi-level graph partition technique, the whole map is recursively divided into submaps in a top-down manner, which finally leads to a binary tree representation. Then all submaps are optimized individually, and merged in a bottom-up manner. By optimizing and merging submaps rather than optimizing the whole map, the proposed method achieves higher efficiency and comparative accuracy compared with existing bundle adjustment methods. Experimental results show that if the sparse structure of the map remains unchanged and the number of cameras in the map is large, the computational cost of the proposed method is proportional to the number of cameras, which is superior to the whole map optimiza...
修改评论