Design of Place Recognition Algorithm Based on VLAD Code and Convolutional Neural Network

Abstract

Visual place recognition is an important and challenging research topic. With the increasing complexity of vision recognition tasks, the traditional image recognition algorithm can not deal with the problem of large-scale illumination change and the interference caused by object occlusion in image. In recent years, deep learning has made many achievements. The main solution proposed based on the deep learning method is to design an end-to-end recognition network, and use pre-labeled place datasets to train the network and finally obtain the classification results of the network. In this paper, a network of place recognition algorithm is proposed which combines deep convolutional features. The sensing field is expanded by adding convolution module, and the extracted convolution features are sent to the local aggregation vector module to obtain the place description vector. The training result of Tokyo Time Machine dataset and Pittsburgh dataset shows that the improved algorithm proposed in this paper has a better recognition recall rate. The generalization performance of the general place retrieval dataset, such as Oxford architecture and Paris architecture, is better than that of the contrast algorithm.