SuperPointVO: A Lightweight Visual Odometry based on CNN Feature Extraction

In this paper, we propose a lightweight stereo visual odometry (SuperPointVO) based on feature extraction of convolutional neural network(CNN). Compared with the traditional indirect method of VO system, our system replace the hand-engineered feature extraction method with a CNN-based method. Based on the feature extraction network SuperPoint, we discard the redundant descriptor information it extracted, and expand the expression ability of the descriptor through NMS and grid sampling, making it more suitable for VO tasks. We build a complete stereo VO system without loop closing around the modified feature extractor. In the experiments, we evaluate the performance of the system on the KITTI dataset, which is close to other state-of-the-art stereo SLAM system. This shows that the accuracy and robustness of feature extraction methods based on deep learning are comparable to, or even better than the traditional methods in VO tasks.