Integrating Human Visual Perception in Scene Understanding by Saliency Visual SLAM

In this study, we propose Salient SLAM, a real-time 3D map generation with saliency information based on human visual characteristics. We used a Fine-Grained technique to generate a saliency map on the spatial domain. Then, we integrated this saliency map to a real-time feature-based Visual SLAM technique. As a result, we can obtain a three-dimensional feature map with saliency information and enable the ability to use this saliency information in robots' autonomous and exploration. Finally, the Salient SLAM map point is leveled into four levels where the higher the level, the higher the score of saliencies in the environment. In the experiments, we investigated and analyzed the characteristics of each saliency level in global feature maps according to the type of objects and visual patterns.