A Downsample Strategy for AGV Life-Long Localization

Abstract

Automated Guided Vehicles (AGVs) have been used in many fields, such as factories, warehouses, etc. They are considered safe, efficient and reliable. Due to the low-cost, light-weight camera and IMU, AGVs based on visual-inertial navigation systems (VINS) have attracted a great deal of interest. Long term localization is essential for AGVs. In real-world applications, this system has to limit resource usage. In this work, we present a sparse strategy of back-end in localization process. We select keyframe according to its information and add it to pose graph maintained in the back-end. To evaluate our approach, we tested our method using a real-world dataset. Our results demonstrate the AGV worked over a long term and was able to detect accurate relative pose. Simultaneously, resource usage in the localization phrase is stable.