Metode Pengenalan Tempat Secara Visual Berbasis Fitur CNN untuk Navigasi Robot di Dalam Gedung

摘要

Place recognition algorithm based-on visual sensor is crucial to be developed especially for an application of indoor robot navigation in which a Ground Positioning System (GPS) is not reliable to be utilized. This research compares the approach of place recognition of using learned-features from a model of Convolutional Neural Network (CNN) against conventional methods, such as Bag of Words (BoW) with SIFT features and Histogram of Oriented Uniform Patterns (HOUP) with its Local Binary Patterns (LBP). This research finding shows that the performance of our approach of using learned-features with transfer learning method from pre-trained CNN AlexNet is better than the conventional methods based-on handcrafted-features such as BoW and HOUP.