Binocular SLAM Based on Learning-based Feature Extraction

The SLAM systems based on the deep learning feature extraction methods usually sacrifice efficiency for accuracy, and there are difficult to meet the system real-time requirements. To deal with such problem, this paper presents a binocular SLAM system that combines deep learning and model compression methods, which includes a local feature computation model built by MobileNetV2 encoder and SuperPoint decoder, and supervises the training of the model through knowledge distillation, and then passes the extracted local feature points and descriptor information to the tracking, local mapping and loop detection threads to realize the complete SLAM system. The experiments test on public datasets and compare with the popular ORB-SLAM2 framework, and the results show that the algorithm in this work can quickly extract local features while meeting the accuracy requirements.