Camera pose estimation method and device

Abstract

The invention discloses a camera pose estimation method and device, and belongs to the field of pose estimation. The method comprises the following steps: acquiring an image and IMU data from a data set; carrying out convolution and pooling operation on the image by using a CNN to obtain feature points and corresponding descriptors of the image, and then carrying out calculation to obtain motion data of the image; processing the IMU data by using LSTM (Long Short Term Memory) to align the motion data with the IMU data; and combining the aligned motion data and IMU data, and sending the combined data to a full connection layer of the CNN for feature fusion and pose estimation to obtain motion features of the camera. The device comprises an acquisition module, a CNN module, an LSTM module and a combination module. According to the method, the problem of camera positioning in VSLAM is solved, the defect that the capability of processing IMU high-frequency data is insufficient in the priorart is overcome, and the calculation precision and the positioning accuracy are improved by adopting a method of combining high-frequency inertial data and low-frequency images.

Description

technical field [0001] The present application relates to the field of pose estimation, in particular to a camera pose estimation method and device. Background technique [0002] VSLAM (Visual simultaneous localization and mapping, vision-based real-time positioning and map construction) refers to the process of calculating one's own position while building an environmental map based on the information of the visual sensor, which can solve the problem of positioning and map construction when moving in an unknown environment , more precisely and quickly. The current mainstream VSLAM framework is mainly affected by real-time, environment, lighting and other conditions. The output of the SLAM (Simultaneous Localization And Mapping, instant positioning and map construction) system includes the positioning of the camera (that is, the motion state of the camera) and the environmental map. For image-based camera localization and VSLAM systems, maps are an important component. Th...

AI Technical Summary

Problems solved by technology

[0003] However, most of the existing camera positioning methods based on deep learning only obtain the camera pose based on pixel and depth data, which easily leads to inaccurate estimation of the camera trajectory and pose.

Images