ATopNet: Robust Visual Localization for AMR Navigation

This article tackled some adverse influences on the localization problem in challenging indoor corridor/aisle environments. We proposed Automatic Topological Localization Network (ATopNet), suitable for autonomous mobile robots (AMRs) based on low-resolution RGB vision. We provided experimental evidence that ATopNet enables AMRs to conduct autonomous routing along indoor corridors with sufficient spatial precision and accuracy that allows AMRs to conduct the following manipulation works. To effectively enhance the performance of the deep Convolutional Neural Network (ConvNet) for visual localization of AMR, we developed automatic training augmentation schemes for the pose, illumination, speed, and occlusion enhancements. We derived the kinematic theories correlating view-frames and observer’s poses, based on which thousands of pose-explicit images can be automatically generated and annotated for ConvNet training. In addition, the strategy of recruiting a conditional Generative Adversarial Network for illumination image generation was proposed. The impact of vehicle speed on localization performance when trained on static images was also tested and discussed. A series of experiments were conducted to verify the efficacy of the proposed methods. We demonstrated ATopNet performing real-time robot localization with a Raspberry Pi4+NCS2. Another prominent feature of ATopNet is that all the procedures are fully automated and require minimal manual operation.