Visual Loop Closure Detection Based on Lightweight Convolutional Neural Network and Product Quantization

Abstract

Mobile robots rely heavily on the creation of the scene map and the positioning in the map in an unknown environment, but no matter what type of map is created, it is inevitably affected by cumulative errors. This presents a huge challenge for loop closure detection technology. Using traditional loop closure detection methods to perform scene recognition is difficult to extract the appearance changes caused by time, weather, or seasonal conditions in the image and deep semantic information, and the speed of extracting image features is slow, which is difficult to meet the real-time performance of robots. Because of the success of deep convolutional neural networks(CNN), it is possible to enrich the information of image features. First of all, this paper uses the pre-trained CNN model SSE-Net to extract the deep visual appearance and semantic features of the image, and obtain the feature description vector. Then, after product quantization(PQ) and encoding, the final pair of candidate frames is quickly searched and matched to obtain the most similar pair of candidate frames and judged as a loop . After the verification of the New collage dataset and the City Center dataset, this algorithm has achieved a good Precision-Recall rate and a faster speed compared with the recently proposed large-scale convolution network VGG16 method and traditional feature extraction methods.