Research On Visual Localization Based On Attention Convolution Neural Network

With the sweeping robot,intelligent express car and other service-oriented products gradually entering people’s life,navigation and positioning technology is also gradually known by people.Visual simultaneous localization and mapping(VSLAM)is an important technology for mobile devices to explore independently in unfamiliar environments,which has important research value.Among them,visual odometry(VO)and loop closing are the core links of visual slam system.The main task of visual odometer is to estimate the camera pose through the relative motion between images,and the estimation result determines the initial positioning accuracy of the whole visual slam system.Loop closing is used to judge whether the camera passes through the same scene during the movement process,which can reduce the cumulative error in the estimation of pose by VO.However,the traditional VO and loop detection methods are mostly based on manually designed feature points,which have some disadvantages,such as time-consuming feature extraction,cumbersome solution steps,poor robustness and so on.(1)Visual odometer based on convolutional neural networks(CNN)and gated recurrent unit(GRU)In view of the cumbersome steps of the traditional visual mileage calculation method based on feature points,a visual mileage calculation method based on convolution neural network and gated loop unit is proposed to estimate the camera pose end-to-end.This method takes the two adjacent image sequences in the data set as the input,extracts the features through the convolution neural network,uses the gated loop unit to learn the timing and establish the motion constraints between the images,reduces the dimension through the full connection layer and outputs the relative pose of the camera with six degrees of freedom.The experiment is completed through Kitti data set,the output pose of the model is visualized and its error is evaluated.The results show that the pose estimation error is small and the trajectory accuracy is high,which verifies the rationality of the visual odometer model.(2)Visual odometer based on attention and convolution neural networkIn order to further improve the pose accuracy of visual odometer model,a visual odometer calculation method based on attention convolution neural network is proposed.The attention mechanism is integrated into the feature extraction module of the visual odometer model to improve the attention of the model to important features,so as to enhance its feature extraction ability,and then improve the overall accuracy of the model.The experiment is completed based on Kitti data set.The ablation experiment verifies the impact of attention mechanism on the performance of the overall model.The accuracy of the overall model is evaluated through comparative experiments.The results show that compared with the feature extraction with only convolution network,integrating attention mechanism can achieve better results,and the algorithm is more accurate than other comparative methods.(3)Loop closing based on convolutional neural network and Net VLADAiming at the low accuracy of the traditional loop closing algorithm based on word bag model,a loop closing algorithm based on convolutional neural network and netvlad is proposed.Convolutional neural network module is improved from VGG structure to extract image features.The netvlad module is improved from vector of locally aggregated descriptors,which is used to cluster the output characteristics of convolutional neural network.Firstly,the features of the image are extracted by the CNN module,then the features are clustered by the netvlad module,finally,the cosine distance between eigenvectors is used to judge whether there is a loop closing.Based on the classical loopback test data set new college and city center,the experiment is completed.The accuracy of the model is evaluated through the visualization of similarity matrix and accuracy recall curve.The results show that the accuracy of loop closing of the algorithm is high.