Neural Network-Based Long-Term Place Recognition from Omni-Images

In robotics perception tasks, visual place recognition has drawn attention as a significant research topic on the grounds of its agile applications without using the global positioning system such as mobile robot navigation, augmented reality, and self-driving vehicles. Owing to the great performance improvement in most computer vision challenges based on deep learning, visual place recognition follows this trend. In this paper, we handle long-term visual place recognition. The long-term visual place recognition can be simplified by substituting it for a conventional supervised classification problem using a convolutional neural network. The proposed network is learned through only a single fisheye-formed illumination-invariant image, captured on Google Street View, for each class. Afterward, sequences of omnidirectional photographs measure how well the network performs. Even though a four-year gap exists between the two datasets, it seems that the proposed network discriminates well against challenges stemming from extreme visual changes.