Optimized Object-aided SLAM for robots Based on Shufflenetv3 and Relocalization

Abstract

Autonomous mobile robots have experienced rapid development over the past few years, and real-time object detection and localization are considered to be their basic capabilities. In order to improve the robot's object detection and relocalization performance, an optimized object SLAM based on Shumenetv3 and relocalization is proposed. Firstly, an enhanced lightweight object detection model Shufflenetv3 is suggested to be used to obtain the object detection of a specified object or multiple objects in the to-be-tested frame, which provides it with the benefits of quick computation and excellent accuracy. Secondly, the accuracy of relocalization is increased by employing the refined target object-aided technique. Finally, to validate the effectiveness of the improved lightweight target extraction, we evaluate the performance of our method utilizing the Pascal VOC 2012 datasets and compare it with YOLOv5 and YOLO-shufflenetv3 in precision and efficiency. Comparing the proposed method to YOLOv5 and YOLOshumenetv3, these quantitative evaluations demonstrate that it performs well with competitive inference times and the most memory-efficient inference. Additionally, we conduct experiments using a public TUM dataset, and the outcomes showed that our approach is far more reliable and consistent to guarantee robot application.