Abstract: In order to solve the problem that the error of the previous frame will be propagated to the next frame when the unmanned vehicle moves for a long time in a complex scene, resulting in the error accumulation of the visual inertia odometry, a multi-state constrained Kalman filter visual inertia odometry algorithm based on key frame loopclosure constraints is proposed. Firstly, the pose of the key frame with fixed time interval is preserved to make full use of the image information and limit the state growth effectively.Then, the loop closure detection is carried out using the bag of words model to determine the key frame where the loop closure occurs, and the observations of loop closure constraintsare add to the feature track for measurement update.Finally, validation analysis is performed in both public datasets and real environments. Experimental results show that compared with MSCKF algorithm, the proposed algorithm can effectively reduce the positioning error and get closer to the real motion trajectory,with higher positioning accuracy and better robustness.

Key words: visual inertia odometer, MSCKF, bag of words model, loopclosure detection, key frame