Visual Place Recognition Based On Convolutional Neural Network

Visual place recognition is a challenging problem in computer vision,and has wide applications in mobile robotics,autonomous driving,etc.Drastically appearance changing in the scene due to illumination and viewpoint will cause perceptual aliasing and perceptual variability,which hinders the long-term application of such algorithms in real environments.In recent years,Convolutional Neural Network(CNN)features have been proved superior to traditional features in visual place recognition.Many existing CNN-based visual place recognition methods directly use the distance of the CNN features and set thresholds to measure the similarity between the two images,which show a poor performance when drastically appearance changing in the scene.Focusing on perceptual aliasing and perceptual variability due to drastically appearance changing in the scene,a novel feature difference map based visual place recognition method is proposed.Firstly,a CNN pretrained on scene-centric dataset is adopted to extract features for perceptually different images of same place and aliased images of different places.Then,a feature difference map is constructed based on the CNN features to represent the difference between the two images.Finally,visual place recognition is regarded as a binary classification task.The feature difference maps are used to train a new CNN classification model for determining whether the two images are from the same place.The innovated results of this thesis are stated as follows:1.Feature difference maps constructed by CNN features are used to represent the difference between the image pairs rather than simply measuring the features from CNN.When drastically appearance changing in the scene,The feature difference map can effectively avoid the impact of perceptual aliasing and perceptual variability and have a higher recognition accuracy.2.Different CNN layers have different robustness to visual appearance and viewpoint.Mid-level features exhibit a robustness against appearance changes,while higher level features are more robust against changes in viewpoint and carry more semantic information.Fusing different CNN features to construct multi-level feature difference map for visual place recognition.3.Visual place recognition is regarded as a binary classification task when drastically appearance changing in the scene.A new classification model is proposed.Compare with existing widely used models,the classification model is more suitable for training feature difference maps and achieves a better recognition performance.