Deep Learning Based Research On Visual Place Recognition

At present,visual place recognition has important application prospects and scientific research value in many relevant science and technology fields,and has become a hot topic in pattern recognition.To exploit a visual place recognition system with better performance that also fits to its application background is significant for improving human life.The key issue for enhancing the performance of visual place recognition system is to design an feature-based image representation method that are more discriminative and more robust to various visual transformations.However,owning to the complexity and diversity of places and the numerous of their application backgrounds,the existing approaches cannot completely meet the needs from the real world,and is still in urgent needs for further researching and exploring.Based on the deep learning approaches,this paper conduct a research and exploration on two difficult visual place recognition tasks which are under two different application backgrounds,and our main works are illustrated as follow:On the basis of extensive reading of domestic and foreign literatures of visual place recognition as well as related fields,this paper reviews the existing approaches of two kind of visual place recognition task—the visual loop closure detection and the cross-view imagery place recognition system.Further,it summarizes the important achievements of the existing methods and the remaining potentials,which lays a foundation for proposing innovative methods in this paper.To handle the visual loop closure detection task,this paper proposes a convolutional neural network named Hybrid-CNN,which based on Squeeze-and-Excitation structure and parallel multi-path convergence compensation strategy,to obtain the improved CNN feature-map.Additionally,we design a feature-map dimension reduction method based on channel-wise downsampling and non-overlapping pooling to accelerate the measurement of pair-wise feature similarity.The classical ResNet is adopted as the base model,and two compensation strategies are conducted to modify the base model to obtain the hybrid-CNN.The designed dimension reduction strategy is implemented to generate the scale-reduced feature-maps,which are applied to achieve the loop-closure detection.The accuracy performance of our approach is verified by comparing to various existing methods on the mainstream open datasets,and the results indicate that the scale-reduced feature-map can remarkably improve the efficiency with negligible compromising on recognition accuracy.To settle cross-view imagery place recognition task,this paper proposes a approach based on our Siam-FCAMNet feature extraction model with a novel exemplar reweighting triplet loss.First,we explore the compensation of attention mechanism to CNN feature representation,and therefore we propose a lightweight attention module,called Mul-FCAM,which based on feature channel attention and spatial context information attention.This module is implemented to modify the ResNet to obtain our ResFCAMNet.Two ResFCAMNets are conducted to construct the Siamese CNN without sharing weights,and further merged with an auxiliary orientation regression branch to form the Siam-FCAMNet to generate the enhanced CNN features applied to that task.Besides,this paper proposes a novel triplet loss which realizes online hard exemplar mining based on exemplar reweighting.It aims to emphasize the positive influence of informative hard exemplars,and suppress the harmful neutralization effect brought by simple triplets,simultaneously.Inspired by the Semi-Hard strategy in FaceNet,we devise a modified logistic regression by adding a rectified distance factor to obtain the correct matching probability of each triplet in the current mini-batch.Then,the information of exemplars are computed according to the rectified correct matching probability to serve as the weights.To restrain the influence of the extreme hard data and the useless simple exemplars,the final assigned weights is pruned by a devised upper and lower bound regulation.Our approach is evaluated by comparing to the current state-of-the-art methods on the mainstream datasets,the results show that our approach can realize an outstanding accuracy performance and achieves state-of-the-art.