Abstract:With the continuous innovation and improvement of sensor technology and SLAM methods, visual SLAM methods have been able to achieve dense reconstruction of the surrounding environment, but when the optimization is poorly initialized, the visual SLAM method suffers from a lack of robustness in the tracking step. Inertial Measurement Unit (IMU) can quickly measure the angular velocity and acceleration of the sensor body. Inspired of this performance, we propose a method for tightly-coupled IMU data with camera data and ensure real-time performance on the GPU. Our method is based on the visual method ElasticFusion. When the global consistency is considered, the sensor velocity, pose, IMU biases are jointly optimized and add inertial information into maps, fully dense surfel-based 3D reconstruction of the environment. Through a large number of experiments, we have found that our vision-inertial navigation SLAM method is significantly better than pure-vision methods in terms of real-time reconstruction, closed-loop capability and map drift correction.

Real-time Indoor Dense 3D Reconstruction of Integration Inertial Measurements Into Surfels