Research On Visual Odometry Method Based On Self-Supervised Deep Learning

With the development of intelligent robots,intelligent robots have been widely used in all walks of life.Visual odometry is the key technology of intelligent robot in the fields of localization and mapping,environmental perception,automatic driving,virtual reality and so on.In recent years,deep learning and convolutional neural network have made great achievements in the field of computer vision,which makes researchers begin to study the visual odometry based on deep learning.The visual odometry method based on geometry traditionally requires high image quality,and the effect is poor when dealing with low texture images.At present,the visual odometry method based on deep learning is divided into supervised and unsupervised ways.The supervised method requires a large amount of labelling data.There are some problems in unsupervised way when estimating the depth map,such as blurred image edge and low accuracy.Aiming at the existing problems in the current methods,this paper proposes and studies a visual odometry method based on self-supervised deep learning.The main work of this paper includes the following four aspects:1.This paper studies and analyzes the current situation of visual slam odometry at home and abroad.Aiming at the fuzzy edge of the depth map estimated in the existing odometry methods based on deep learning,we proposed a visual odometry method based on self-supervised deep learning.It includes two networks,namely depth estimation network and pose estimation network.They complete two tasks of depth estimation and pose estimation at the same time.In the odometry network structure,the residual network Res Net50 is taken as the backbone network structure,and a CBAM module is added behind each residual block.The CBAM module can enhance the features conducive to training and reduce the redundancy of invalid features.Experiments indicated that compared with the existing odometry methods based on deep learning,the method proposed in this paper can improve the edge ambiguity of depth map and have a certain improvement the accuracy of pose.2.Aiming at the situation that the odometry method of supervised deep learning needs a large number of labeled data,but in practice,the production cost of labeled datasets is high.In this paper,the self-supervised method is used in visual odometry model training,the method takes the monocular color map as the input of the model,reconstructs a new target image through the odometry network,and then takes the next frame of image data as the supervision information to establish a data relationship with the reconstructed target image,so as to supervise the training of the network model and transform the self-supervision problem into an image reconstruction problem.In order to improve the effect of depth estimation,in the design of loss function,in addition to the main image reconstruction loss,it also increases the loss of image similarity and smoothing loss.Experiments indicated that the self-supervised deep learning method is feasible in dealing with some specific visual problems,and provides ideas for other visual tasks.3.Aiming at the problem of limited computing resources of embedded intelligent mobile devices,a lightweight odometry network model based on Mobilenetv2 network is proposed.The lightweight network model uses the reverse residual structure,which reduces the complexity of the network model and can effectively reduce the parameters of the odometry network model.Experiments indicated that the accuracy of the lightweight odometry model proposed in this paper will not decline significantly when the network parameters are greatly reduced.4.Finally,a visual odometry system based on self-supervised deep learning is designed.The system supports two input modes: offline video stream and real-time image capture.The visual odometry system is deployed to the intelligent mobile robot with limited computing resources,and only the reasoning of odometry model is carried out at the robot terminal.In the real scene,the intelligent mobile robot is equipped with a camera sensor to collect image data in real time,estimate the scene depth and robot pose,and realize the end-to-end visual odometry method.