CEP-SLAM Based on RGB-D Data Coupling Errors Processing

Aiming at the issues of coupling errors in RGB-D data, incorrect extraction of edge-point in current methods of feature extraction and poor tracking stability of the constant speed motion model in VSLAM using RGB-D cameras, CEP-SLAM algorithm was proposed based on ORB-SLAM2 framework. The algorithm used a constant acceleration motion model to set the initial pose of the tracked frame; the optimized pose was used to calculate the inter-frame visual odometry and update the constant acceleration motion model. The pose deviation was estimated by combining the constant acceleration motion model and the acquisition time difference of RGB image and depth image. The epipolar geometry constraint was constructed based on the pose deviation and the dichotomy method was used to find the position of the feature-point in the corresponding pixel point of the depth image, and the depth of the feature-point was adjusted to alleviate the impact of RGB-D data coupling errors on VSLAM; a keyframe edge-point culling algorithm based on joint method was proposed. The bad edge-point in the inserted key-frame were judged and culled by using the neighborhood information of feature-point in the depth image. The CEP-SLAM algorithm proposed in this paper was used to conduct experiment on TUM public dataset. The experiment results show that the proposed algorithm can better cull bad edgepoint, and has better robustness, tracking stability and higher positioning accuracy compared with classical algorithms.