Detection of moving features using IMU-camera without knowing both the initial conditions and gravity direction

Detecting moving features relative to ground in the images of a moving camera is important for mobile robot localization in practice. This problem is particularly difficult if the initial conditions of the camera are unknown. In this paper, we propose a moving feature detection method by using a calibrated IMU-camera in a dynamic environment. The proposed method is able to separate static and dynamic features without knowing the IMU-camera initial conditions, as well as the gravity direction. In the method, an estimator initialization algorithm is implemented first to estimate the moving velocity and 3D positions of the feature points, and the gravity direction. Then, a recursive moving object detection algorithm is designed to classify the static and dynamic features based on feature re-projection. The simulation results show that the moving features can be grouped effectively, and the remaining static feature points can be used for camera pose and velocity estimation in a real scale to the ground.