A Visual Inertial Odometer Method Based on Dynamic Object Detection

This paper proposes a visual inertial odometry algorithm based on dynamic object detection. The algorithm begins by enhancing the Mask R-CNN instance segmentation method through the introduction of a cascaded image pyramid to improve its recognition capabilities. Subsequently, a dynamic object detection scheme combining optical flow and epipolar geometry constraints is designed to achieve rapid detection of dynamic objects. Experimental results demonstrate that the algorithm enhances the accuracy and stability of SLAM systems in dynamic environments using visual sensors.