SupAtten-SLAM: SuperPoint Variant Network Visual SLAM Based on Multiattention Mechanism

Simultaneous localization and mapping (SLAM) system is the basis of intelligent mobile robot working in unknown environment. Traditional visual SLAM methods may encounter the problems of inaccurate positioning and unstable matching when dealing with complex scenes such as rotation, illumination change or lack of texture. To overcome these challenges, we propose a SuperPoint variant network based on the multi-attention mechanism. In the implementation process, we first encode the features of the input images through the backbone network, and then share the encoded information through the feature detection subnetwork and the feature descriptor subnetwork. In this way, we propose a novel convolutional neural network and integrate it into a visual SLAM system for testing. The experimental results show that compared with the existing visual SLAM system, the proposed method shows excellent performance in a variety of complex environments and surpasses the traditional visual SLAM in overall performance.