Enhancing Visual Navigation Performance by Prior Pose-Guided Active Feature Points Distribution

Feature points are the principal measurements employed in visual navigation, and their tracking qualities significantly affect the localization accuracy of visual-based navigation systems. However, the current feature extraction methods do not take into account the expected tracking quality of the features in the subsequent steps, which may result in less tracking length or limited parallax of the features. Such expected feature tracking quality can be inferred using the upcoming pose information of the carrier, which is a known factor for robots and autonomous vehicles through path planning. Based on the typical grid-based feature distribution method, this paper proposes a prior pose-guided active feature points distribution method to redistribute the feature number in different image regions. In the proposed method, the prior poses of the carrier and the three-dimensional (3D) environment that represented by the sparse triangulated features are utilized to predict the future tracking of the features. The feature numbers for each image region are reassigned according to the predicted tracking parallax. The proposed method is implemented in a multi-state constraint Kalman filter (MSCKF)-based visual-inertial navigation system and evaluated with the public dataset and our private robot dataset. The experiment results indicate that the proposed method can effectively enhance the average tracking parallax of the features and improve navigation performance significantly. The robustness tests also confirm that intentionally introducing certain noises in the prior poses does not hinder the proposed method. Depite these perturbations, the proposed method still demonstrates superior navigation accuracy than the conventional grid-based method.