Suman Deb

Perceiving and understanding dynamic scenes by a robot vision system requires crucial spatio-temporal knowledge and geometry to enable long-term autonomy. Existing Dynamic SLAM (Simultaneous Localization And Mapping) addresses definite portions of robot pose estimation to make it accurate. On the contrary scene understanding and rational integration with position is hard to find. In this context, a comprehensive realisation of a variable scene shared with other dynamic agents is still a challenge to execute within a shifting drift. This challenge is even increased when the environment is with variable luminance. To address these challenges, we propose a spatial-geometric SLAM (Spatio SLAM) solution that unifies the exiting connotations of long and short-term dynamics that constructs a real-time dense spatial-geometric map. The geometric representation of the scene gives the autonomous agent a prior map to factorize the navigation framework. In this work an end-to-end DA-SSOD (Domain Adaptive Single Stage Object Detection) deep network is introduced into the object detection module. This DA-SSOD introduction experimentally confirms the accurate prediction of obstacles in complex indoor surroundings with variable luminance. Geometric maps built by Spatio SLAM (as the real-time reconstruction of a 3D scene over an instance), results in 86.6%precision in static and 87.4%precision in varying environment (with dynamic obstacles). SOTA (State Of The Art) analysis presents the superiority of Spatio SLAM on other SLAM techniques implemented across multiple indoor structures.