Place Recognition with DAG-CNN

This paper presents the development of a convolutional neural network with a directed acyclic graph architecture (DAG-CNN) focused on the recognition of places. The network is focused on identifying six types of rooms in various houses. For this purpose, five houses have been built in a virtual environment from which the training and validation database has been obtained through an on-site panning camera. In order to select the number of filters required for the proposed architecture, the internal behavior of each training has been verified through neuron activation heat maps in order to reduce the learning repetitions of little relevant objects or the characteristics of the scene as much as possible, obtaining a network capable of recognizing 96.5% of the individual images from room sequence photographs and 100% individual recognition of each room (complete sequence). Thus, the capacity and the robustness of the selected architecture for recognizing indoor places are demonstrated.