It is highly challenging to ensure accuracy and robustness of simultaneous localization and mapping (SLAM) algorithms with rapid camera movements and drastic viewpoint changes. In these scenarios, feature extraction, feature matching, and tracking become very difficult. To address this issue, visual and visual‐inertial SLAM based on enhanced deep learning features and motion smoothness constraints is proposed in this work. A lightweight convolutional neural network is integrated into geometry‐based visual and visual‐inertial SLAM to enhance the quality of extracted features. Next, feature matching and tracking are respectively improved. The rotation angle information is computed and integrated for the extracted deep learning features. Based on the statistical results of rotation angle information, the feature matching point pairs with low accuracy are eliminated. Motion smoothness constraints are introduced to the tracking process using a grid‐based motion statistics method, thereby enhancing the performance of SLAM tracking. The performance of the proposed SLAM is evaluated through dataset experiments and robot experiments in real‐world scenarios. The outcomes demonstrate favorable performance of the proposed SLAM compared to other state‐of‐the‐art algorithms.
Jiang et al. (Thu,) studied this question.