Attention-SLAM: A Visual Monocular SLAM Learning From Human Gaze

This paper proposes a novel simultaneous localization and mapping (SLAM) approach, namely Attention-SLAM, which simulates human navigation mode by combining visual saliency model (SalNavNet) with traditional monocular SLAM. Firstly SalNavNet is proposed. In SalNavNet, we introduce correlation module propose an adaptive Exponential Moving Average (EMA) module. These modules mitigate the center bias, most current models have. idea enables maps generated to pay more attention same salient object. An open-source SLAM dataset Salient-Euroc published, it consists of Euroc corresponding maps. Moreover, new optimization method called Weighted Bundle Adjustment (Weighted BA) in Attention-SLAM. Most methods treat all features extracted from images as equal importance during process. weighted BA, feature points regions have greater importance. Comprehensive test results prove that Attention-SLAM outperforms benchmarks such Direct Sparse Odometry (DSO), ORB-SLAM, Salient DSO 7 11 cases. The cases are indoor scenes, varying brightness, speed, image distortion. Compared our improves accuracy 4% efficiency 6.5% on average.