Towards Loop Closure Detection for SLAM Applications Using Bag of Visual Features: Experiments and Simulation

This paper presents a new approach to exploring sparse and binary convolutional filters in traditional Convolutional Neural Networks (CNN). Recent advances in the integration of Deep Learning architectures, particularly in mobile autonomous robotics applications, have motivated several researches to overcome the challenges related to the limitations of computational resources. One of the biggest challenges in the area, is the development of applications to address the Loop Closure Detection problem in Simultaneous Localization and Mapping (SLAM) systems. For such application, it is necessary to use exhaustive computational power. Nevertheless, resource optimization of Convolutional Neural Network models enhances the capability of integration. Therefore, we propose the reformulation of convolutional layers through Local Binary Descriptors (LBD) to achieve this kind of optimization of CNN's resources. This paper discusses the evaluation of a Bag of Visual Features (BoVF) approach, extracting features through local descriptors (e.g., SIFT, SURF, KAZE), and local binary descriptors (e.g., BRIEF, ORB, BRISK, AKAZE, FREAK). The descriptors were evaluated in the recognition and classification steps using six visual datasets (i.e., MNIST, JAFFE, Extended CK+, FEI, CIFAR-10, and FER-2013) through a Multilayer Perceptron (MLP) classifier. Experimentally, we demonstrated the feasibility of producing promising results by combining BoVF with MLP classifier. Additionally, we can assume that the computed descriptors generated by a Local Binary Descriptor alongside the proposed hybrid DNN (Deep Neural Network) architecture can satisfactorily accomplish the results for the optimization of a CNN's resources applied to the Loop Closure Detection problem.