Research On Visual Place Recognition Method Based On Convolutional Neural Network

With the development of science and technology,mobile robots play an increasingly important role in human production and life.Visual place recognition is the basis for mobile robots to realize autonomous positioning and navigation.Its task is to judge whether the newly arrived place is consistent with the previously visited place.In the process of visual place recognition,it will be affected by the surrounding environment lighting,weather and other conditions,and there may also be interference from the camera's viewing angle change,resulting in performance that is not as expected when using traditional manual features for place recognition.In recent years,image feature extraction technology based on convolutional neural networks has been widely used in the field of computer vision,which provides new ideas for improving the performance of place recognition.By analyzing the performance of features extracted by convolutional neural network in the process of visual place recognition,the following research work is mainly carried out:(1)A method of visual place recognition based on multi-layer attention mechanism is proposed.Firstly,in view of the problem that the features extracted by the convolutional neural network are sensitive to the change of the image perspective,the context attention mechanism is used to learn the hidden information of the image,so as to assign weights to different regions of the convolutional features.Secondly,in view of the problem that different convolutional layers contain different levels of image information,a multi-layer attention mechanism is established to fuse multi-scale convolutional feature maps.Finally,the cosine distance is used to calculate the similarity between image descriptors.Place recognition experiments were conducted on the public datasets St.Lucia,Synthesized Nordland and Gardens Point.The improved method with multi-layer attention has a significant improvement effect compared with the basic convolutional neural network.It can be verified by the visual analysis of the heat map that the method can effectively identify the region of interest in the image and enhance the robustness of the convolution feature.(2)A method of visual place recognition based on cross-layer attention mechanism is proposed.Firstly,in view of the problem that the superposition of multiple convolution kernels is likely to cause the loss of the image edge receptive field,the symmetrical separation convolution is used to construct a spatial attention mechanism to identify the region of interest in the image.Secondly,in view of the problem of information redundancy that is likely to cause information redundancy when learning different layers of convolution features multiple times,the spatial attention mask learned from the high-level convolution features is multiplied by the low-level convolution features.Finally,combine the local feature aggregation descriptor method to construct a compact global feature descriptor vector.Experiments were conducted on the public datasets St.Lucia,Synthesized Nordland and Gardens Point.The improved method of cross-layer attention has higher recognition accuracy in complex environments.It can be verified by the heat map visualization experiment that the method can identify the stable and static regions of the image,and it shows stronger robustness than the basic convolutional neural network.(3)Campus environment datasets under different environmental conditions were collected,and place recognition experiments and heat map visualization experiments were conducted.Experiments show that compared with the basic convolutional neural network algorithm,the algorithm proposed in this paper has better recognition performance and good generalization.