Fusion of Visual Odometry Information for Enhanced Camera Pose Estimation

Camera pose estimation involves determining the camera's position coordinates and angular rotations around three axes, describing its orientation and location relative to a given scene. While neural networks have made significant strides in camera pose estimation, they are susceptible to issues like motion blur in the scene. This paper addresses the challenge of motion blur caused by camera movement in camera pose estimation by leveraging information from visual odometry, which provides camera motion and pose details. A visual odometry network is designed, and trajectory drift in the network is mitigated by incorporating Long Short-Term Memory (LSTM). The visual odometry information is fused into the camera pose network to enhance pose accuracy. The proposed method demonstrates substantial improvements compared to other approaches.