Comparative Evaluation of Feature Descriptors Through Bag of Visual Features with Multilayer Perceptron on Embedded GPU System

Abstract

Loop Closure Detection (LCD) is one of the most significant steps for a Simultaneous Localization and Mapping (SLAM) system of an autonomous robot. Many of the state-of-the-art techniques for the LCD problem are based on handcrafted resources and Bag of Words (BoW). Even with advances in Machine Learning (ML) area, Deep Learning (DL) methods and Convolutional Neural Network (CNN) are not fully explored in the LCD problem context. We addressed the LCD problem proposing a hybrid Deep Neural Network (DNN) architecture for deployment on an embedded GPU system: NVIDIA's Jetson Nano. The main idea of this work consists of reformulating convolutional filters of a CNN through Local Binary Descriptors to produce a sparse model as an efficient alternative to traditional CNNs. This paper discusses the evaluation of the Bag of Visual Features (BoVF) approach, extracting features through Local Feature Descriptors such as SIFT, SURF, and KAZE and Local Binary Descriptors such as BRIEF, ORB, BRISK, AKAZE, and FREAK were evaluated for the recognition and classification steps. For this task, we used six visual datasets (MNIST, JAFFE, Extended CK+, FEI, CIFAR-10, and FER-2013) through Multilayer Perceptron (MLP) classifier. Experimentally, we demonstrated the feasibility to produce promising results combining BoVf with MLP classifier. Therefore, empirically, we can assume that the computed descriptors generated by a Local Binary Descriptor alongside a hybrid DNN architecture, can accomplish satisfactory results addressed in future work in the Feature Detection step, and reformulated in the Feature Extraction step into convolutional filters of CNN architecture of the proposed system.