A real-time visual-inertial mapping and localization method by fusing unstable GPS

This paper presents a novel method which fuse visual, IMU and GPS tightly to realize high-precision real-time localization and mapping simultaneously (SLAM). Our method is based on the bundle adjustment (BA). The confidence of the GPS signal is used to determine the window size in the local mapping thread and judge whether the keyframe is reliable. The long-term unreliable keyframe linking with large uncertainty of GPS which called GPS-restricted or GPS-denied situation will cause the drift when mapping. To eliminate the drift, in contrast to use the closed-loop detection and global optimization which will increase the computational burden extremely with the size of the map enlarged, a semi-global optimization method is proposed to relieve the burden, which make the localization estimated by this method possible to be used to navigate for unmanned vehicles. In our method, the confidence of the GPS signal is significantly important, however, the covariance supplied by the GPS receiver may not be trustworthy sometimes, which cause some unnecessary mistake when optimizing, thus a semi-supervised clustering method taking the information of GPS and IMU into account synthetically is introduced to get that confidence more robustly.