中国科学院自动化研究所机构知识库

With the increasing popularity of mobile robots, the demand for indoor localization of robots is growing as well. People are in the hope that mobile robots can achieve accurate localization and navigation indoors, and then perform a series of tasks. Indoor localization has a wide range of applications in factories, homes, shopping malls, etc. For the sake of achieving indoor accurate localization, the visual localization technology is a feasible solution, especially to estimate the camera's six-degree-of-freedom position and orientation in the world coordinate system through a single image shot by the camera, which is easy to operate. However, due to the complex indoor environmental changes, a single image contains less information in comparison with a video. How to improve the accuracy, robustness and efficiency of the localization system based on a single image is still a challenging work. In this paper, a series of researches on the visual location problems of the indoor complex scene based on a single image are carried out. The main contributions are as follows:

1.A medium-sized indoor complex scene test dataset is constructed. In view of the problems that the size of the existing public dataset is too small and the test sample has no ground truth value, this paper uses the laser-vision SLAM scanning equipment Navvis to scan the total area of 8000 square meters of three floors of Wanda Plaza in Shijingshan, Beijing, so as to obtain the 1567 mixed panoramic pictures and the depth images.  In the meantime, the full-stop electronic rangefinder is used to assist the localization in improving the accuracy of the 3D point cloud model. In this paper, 36 perspective views of different perspectives are generated from each panorama by the method of perspective synthesis. The three-dimensional point coordinates corresponding to each pixel of the perspective view are generated by the three-dimensional point cloud model, thus a three-dimensional scene database containing 56412 images is constructed. In addition, 4701 images of the scene are collected by three different types of mobile phones as the test dataset. The internal parameters of the mobile phone are individually calibrated by the two-dimensional checkerboard method. The external parameters of the mobile phone and Navvis device are estimated by EPnP in the test dataset image, so as to obtain the ground truth pose of the test image.

2.A set of visual localization flow based on the image retrieval is designed. In view of the deficiencies of low efficiency and low accuracy of the current visual localization method, this system first compares three mainstream image retrieval algorithms of “BOW”，“Disloc”，and “Inloc”, compare their advantages and disadvantages, and selected BOW model as the final image retrieval scheme. In order to further improve the efficiency of localization, GPU is introduced to accelerate the extraction of the features of RootSIFT on the basis of BOW similarity retrieval. In accordance with the results of RootSIFT feature matching, the retrieval results are reordered. The top-10 most similar images and their corresponding feature matching results are input into the Perspective-N-Point localization algorithm to estimate the query image pose. The experiment results show that the efficiency of this method is more than 10 times higher than that of the traditional algorithm, with the characteristics of high precision, high efficiency and strong robustness.

3.A complete set of indoor vision localization system based on Client-Server architecture is realized. The localization time of this system is 200ms, the localization accuracy is 6cm, the angle accuracy is 0.32 °, and the localization success rate is 91.6%, which meets the needs of practical application. The system includes the following modules: the client module, which is responsible for completing the task of the user's taking the scene image, compressing and uploading it to the server, and receiving the returned localization results; the server module, which is responsible for loading the 3D scene model reconstructed offline, carrying out the  similarity retrieval and the pose estimation for the received client images, and sending the estimated pose to the client.