A Visual SLAM Model Based on Lightweight SuperPoint and Depth Metric Learning

In the tasks of V-SLAM, 3D reconstruction, and SFM, the extraction of image feature points and the calculation of descriptors are very important. The robustness and accuracy of the above algorithms can be significantly improved by better reflecting the feature points of image information and more specific descriptors. In this paper, the SuperPoint network with high robustness and good performance is selected as the feature point extraction algorithm. Select the geometric corresponding network algorithm as extraction descriptor, and finally extract the network model of both script and feature. To solve the problem of large amounts of calculation and parameters, use the Depthwise separable convolution to replace the ordinary convolution, and change the way of down-sampling and the number of convolution layers. Experiments show that the SuperPoint network can only run at 5-10 Hz frequency in i7-9700 and GTX1650 configurations when combined with the ORB-SLAM2 system directly. The improved network model can run with CPU only and keep the frequency above 25 Hz, which is more robust and accurate than the ORB feature point.