A Keyframe Optimization Method Based on Information Entropy

In response to the issue that the VINS-Fusion algorithm primarily relies on heuristic strategies based on parallax magnitude and feature tracking quality during keyframe selection, making it difficult to effectively measure global map uncertainty and potentially leading to insufficient lo-calization accuracy, this paper proposes a keyframe selection optimization method based on in-formation entropy theory. First, by using a quadrotor UAV as the hardware platform, the key processes of front-end feature tracking and back-end sliding-window optimization in VINS-Fusion are systematically analyzed. Then, during the keyframe insertion stage, an in-for-mation entropy metric is introduced to accurately characterize the uncertainties of map points and pose estimates. By setting an entropy threshold, the timing of keyframe insertion is adap-tively optimized, effectively reducing the global cumulative error. Finally, the proposed method is validated on the open-source EuRoC dataset and comprehensively compared with the original VINS-Fusion algorithm. Experimental results demonstrate that incorporating infor-mation en-tropy into the keyframe selection strategy significantly reduces redundant keyframes, decreases map-point uncertainty to some extent, enhances global trajectory estimation accuracy, and has only a minimal impact on the real-time performance of the algorithm.