Enhancing AdultSize Humanoid Localization Accuracy: A Vision-based aMCL Leveraging Object Detection Model and Hungarian Algorithm

Abstract

The robot's decision-making is an essential component of the autonomous robot. To fulfill this component, the accurate position estimation of the robot is a fundamental prerequisite. Localization determines the robot's relative position within the map environment, and existing mobile robots have been extensively studied. However, localization is still challenging for humanoids because the bipeds' movement is not as stable as mobile robots, and camera view frames oscillate from side to side. Developing localization with high accuracy under the above-limited constraints is necessary. This paper proposes an estimation of a 1.3m tall humanoid robot's position with an adaptation of vision-based Augmented Monte-Carlo localization (aMCL) in the soccer field. Several approaches were also applied to enhance the localization performance. First, a deep learning-based object detection model was selected to identify predefined landmarks. In addition, the data association process was improved from the nearest neighbor matching algorithm to the Hungarian algorithm. This method enhanced data association performance, and the robot's position was successfully estimated in real-time.