Object Aware RGBD SLAM in Dynamic Environments

Simultaneous Localization and Mapping (SLAM) is a probabilistic approach that builds the map of an unknown environment and estimate the position of the agent within that map at the same time. SLAM has three assumptions to be able to work properly; the environment must be suitable for data acquisition (illumination, textured, etc.), the data that is received from the sensors must be overlapping partially in consecutive frames and the static environment must be dominated over the dynamic environment.The last autonomous robotics applications have a lot of impacts to our lifes. They already started to work in the indoor and out-door environments (warehouse, caffe, street etc). Most of the use cases of the robots can provide the first two assumptions which is mentioned in the last paragraph. But the third assumption which is related with dynamic objects can be easily violated when the use of robots in dynamic environments with moving objects.In this work, we focused on the decreasing the negative impact of dynamic objects to SLAM accuracy. For this, we developed RGBD SLAM solution that can live with the dynamic objects. The main workflow of the solution assumes visual features that are obtained from dynamic objects as outliers. We used open source version of the ORB-SLAM3 as a base software for RGBD SLAM. We integrated YOLOv3 model to our solution to detect the dynamic objects in RGB frames. We selected ORB-SLAM3 and YOLOv3 applications for success of the algorithms. Also they can work at high frequency which means that they can work in real life applications.In this work, we observed 82% improvement on average in the TUM “wallking” datasets. By elimination of the outliers, the accuracy of the SLAM algorithm is increased. Also the robustness of the system against to the dynamic objects is increased.