Deep Learning-based Algorithms For Visual Odometry

With the development of computer vision and the emergence of various high-performance vision sensors,visual odometry has been widely used in robots,autonomous vehicles,and other fields.Monocular visual odometry is the process of determining the position and orientation of a camera by analyzing a sequence of input images from a monocular camera and has the advantages of being low cost,simple to configure and computationally efficient.monocular visual odometry suffers from scale ambiguity due to the lack of depth information in the monocular camera.Deep Neural Networks(DNN)have been widely used in computer vision tasks,so it is important to study DNN-based monocular visual odometry.Firstly,a lightweight algorithm for local feature detection and description of images is designed to address the problems of large network model parameters and high ha rdware resource consumption of DNN-based feature extraction and description algorithms.In order to obtain multi-scale features of images,a backbone network based on dilated convolution and basic residual blocks is designed;the combination of attention m echanism and jump connection enables the fusion of low-level semantic information with high-level semantic information of the network while extracting better features.The experimental results show that with a model size of only 1MB and much smaller than similar methods,the mean matching accuracy(MMA)values are improved by 2.3%,2.4%,1.9%,1.4% and 1.1% respectively on the HPatches dataset compared to the results of the R2D2 algorithm when the threshold values are in the range of 2to 6 pixels.Secondly,the proposed feature detection and description network were integrated with the geometry-based odometry method to construct a hybrid method visual odometry and verify the algorithm in comparison on sequences 09 and 10 of the KITI Odometry dataset.Compared to the VISO2 algorithm,the proposed hybrid VO algorithm has a 55% and 53% reduction in translation error respectively.Finally,an end-to-end visual odometry calculation method is proposed to address the scale ambiguity problem in monocular visual odometry.A depth estimation network and a bit-pose estimation network are jointly trained in an unsupervised framework,and they are constrained to predict scale-consistent results.The algorithm is compared with the current mainstream endto-end VO and geometry-based algorithms such as ORB-SLAM2 to demonstrate the effectiveness of the algorithm.