Research On Visual Inertial Odometry Algorithm Based On Deep Learning

Simultaneous localization and mapping(SLAM)technology plays a key role in the autonomous mobile navigation of robots.Its front-end odometer is the core part of the positioning function.In some cases,the lighting conditions in the scene are poor and the collected images are blurred,which will lead to poor positioning effect of the visual odometry.Traditional visual-inertial odometry requires accurate sensor calibration and complex modeling to achieve state estimation.The deep learning method relies on a large number of data sets containing real values.In response to the above problems,this thesis designs an unsupervised visual odometry algorithm based on deep learning and then optimizes the network.The model can fuse visual information and inertial information to achieve pose estimation,and does not require complex calibration modeling or a large number of ground truth datasets.The model has high positioning accuracy,and positioning can still be achieved in the case of sensor degradation.The main work is as follows:1.The thesis designs an unsupervised visual inertial odometry method based on deep learning.The algorithm uses optical flow network encoder image features and uses transfer learning to optimize the initial performance of the network and improve training efficiency.The extraction of inertial features is realized through the long short-term memory network,and then the extracted image and inertial features are completed through the pose network to complete the pose fitting.The thesis trains the network with an unsupervised method by building geometrically consistent and photometrically consistent models.The accuracy of the network was analyzed qualitatively and quantitatively in the KITTI dataset,and its average rotation error reached 0.92%,which was lower than other comparative networks.It is tested that the network still has a good positioning effect in the case of sensor degradation,and cross-dataset verification is carried out on the Malaga dataset.2.On the basis of the unsupervised visual-inertial odometry method,the pose optimization of the visual-inertial odometry method is realized through data fusion strategy and network optimization.The self-attention mechanism and the cross-attention mechanism are used to guide the screening of image features and inertial features,and two different types of data features are reasonably fused to complete the pose calculation.The training part of the network is improved by optimizing the loss function and inter-frame optimization.First,a smoothness loss function and a geometric consistency loss function are introduced,and the change of the training loss curve after the improvement of loss function is analyzed.Secondly,the training of the network is optimized through the relationship between frames.In the process of training the network,three consecutive frames of data are sent to the network,and the network training effect is enhanced by calculating the loss of the two frames before and after the image relative to the original image.Finally,it is proved by experiments that the proposed strategy can improve the network positioning accuracyIn summary,this thesis designs an unsupervised visual inertial odometry method based on deep learning methods,and optimizes the network through different strategies.Then the thesis compares and analyzes the effect of the model through experiments on the KITTI and Malaga datasets.