Transfer Learning for Humanoid Robot Appearance-Based Localization in a Visual Map

Autonomous robot visual navigation is a fundamental locomotion task based on extracting relevant features from images taken the surrounded environment to control an independent displacement. In navigation, use of known map helps obtain accurate localization, but in absence this map, guided or free exploration pathway must be executed sequence representing map. This paper presents appearance-based localization method and end-to-end Convolutional Neural Network (CNN). The CNN initialized via transfer learning (trained using ImageNet dataset), evaluating four state-of-the-art architectures: VGG16, ResNet50, InceptionV3, Xception. A typical pipeline for includes changing last layer adapt number neurons according custom classes. work, dense layers after convolutional pooling were substituted by Global Average Pooling (GAP) layer, which parameter-free. Additionally, L 2 -norm constraint was added GAP feature descriptors, restricting lying fixed radius hypersphere. These different pre-trained configurations analyzed compared two maps found CIMAT-NAO datasets consisting 187 94 images, respectively. For tasks, set 278 available each numerical results proved that integrating training pipeline, performance boosted. Specifically, VGG16 Xception networks achieved best results, reaching top-3 accuracy 90.70% 93.62% dataset, respectively, overcoming referenced approaches hand-crafted extractors.