Online Self-Calibration Initialization for Multi-Camera Visual-Inertial SLAM.
Yong Liu,Fan Wang,Wen Zhnag,Chaofan Zhang,Yingwei Xia
DOI: https://doi.org/10.1109/robio.2018.8665294
2018-01-01
Abstract:Multi-camera SLAM has become more popular in robotics, as capable of maximizing perceptual awareness of the environment. Since the multiple sensing modalities offer complementary characteristics that can estimate more precise rigid body movement trajectories, multi-camera visual-inertial SLAM becomes a trend of development. Robust and accurate initialization is considered the core capability for multi-camera visual-inertial SLAM. In the present paper, an online self-calibration visual-inertial initialization method is presented for multi-camera visual-inertial SLAM. Some improvements are made to introduce a monocular visual-inertial initialization method to the multi-camera visual-inertial SLAM system. In particular, a hardware synchronization mechanism is added, a self-calibration method is introduced, and an IMU initialization method is achieved on the basis of multi-camera measurement. First and foremost, a hardware synchronization trigger method with exposure compensation is used, ensuring more accurate time synchronization between multi-camera and IMU, and gaining more reliable visual-inertial measurements. Secondly, a state-of-the-art online self-calibration method is introduced, estimating the camera-IMU extrinsic parameters automatically without extra prerequisites. Besides, the external parameters calibration terminate criterion is followed to launch the initialization procedure. For the initialization procedure, the rich visual measurements from multi-camera are performed to provide priors for the IMU initialization. Next, a novel improvement using the pose of the whole multi-camera system is proposed to gain IMU initialization parameters, which include visual scale, gravity, velocity, biases of gyroscope and accelerometer. To validate and evaluate the presented initialization method, the parameters yielded by initialization are comparatively evaluated, and our method is evaluated in challenging real-world environment.