CNN-based Visual Place Recognition In Changing Environments

Prompted by applications in unmanned aerial vehicle,autonomous driving,virtual reality,visual place recognition has received a significant amout of attention in computer vision and robotics communities in the past few years.Visual place recognition dominated by conventional hand-crafted feature-based techniques such as SIFT,ORB,failed where the problem of changing environmental conditions has come to the fore.Inspired by the astounding performances of CNN-based features in tasks such as image classification,face recognition,a series of CNN-based techniques have been adopted to address this problem.First,a simple visual place recognition pipeline based on a Bag of Convolution Features(BCF)is proposed,which is implemented by encoding CNN-based features using BoW aggregation.Feature-mapping image,which allows directly mapping from regions of the original image to a visual word,is also introduced to perform a fast spatial reranking.Faster R-CNN,with the advantage of object proposals,is also employed in a visual place recognition pipeline.To develop an end-to-end trainable CNN architecture for visual place recognition tasks,a novel network architecture,BoWNet,is designed to mimic the BoW model.Then BoW-CNN is developed by plugging BowNet into any CNN architecture such as AlexNet,VGG16.And an architecture of BR-CNN is also developed with BoWNet and Faster R-CNN.We also develop a training procedure based on a triplet ranking loss.Finally,it's showed that the proposed CNN-based visual place recognition techniques significantly outperforms hand-crafted feature-based techniques on four challenging place recognition benchmarks.