Target-Driven Navigation of ROS Robot with Object Detection using Deep Learning

Target-driven navigation of robot has the task of navigating in an environment to reach a target specified by the user. Mapping, localization and planning are the major challenges in accomplishing the task of target driven robots. To overcome these challenges, this paper proposes a system integrated with Hector SLAM, Adaptive Monte Carlo localization (AMCL), A* and Dynamic Window Approach (DWA). In order to make robot able to navigate through complex environments, surrounding map has to be created. Simultaneous Localization and Mapping (SLAM) is a widely used technique for creating map of an unknown environment and localizing the robot at the same time. Once a map is created, the problem of localization of robots in the map arises. This issue of localization is solved in this system by integrating AMCL. The map created by SLAM algorithm is utilized by path planning algorithm to reach the final coordinates specified by the user. The proposed system uses A* algorithm for global path planning. But global path planning algorithm alone cannot handle new or dynamic obstacles in the path of navigation of robots. To deal with such obstacles, a local path planning algorithm called Dynamic Window Approach is also used in this system. In applications such as indoor service robots, rescue robots etc., the use of object detection algorithms enhance the performance of robots in fulfilling the task. A technique called ensemble method, which combines the results of several models is utilized in this system for object detection and identification. The combined use of Hector SLAM, AMCL, A* and DWA along with ensemble object detection algorithm improves the performance of target-driven navigation of robots. This paper studies the simulation of the proposed system in an open-source framework called Robot Operating system (ROS).