Research On Visual Odometry Technology Based On Deep Learning

VO is a challenging open problem in vSLAM system.Its main task is to estimate its own(sensor or moving vehicle)posture changes based on the image data returned by the visual sensor(usually attached to a moving vehicle,such as a car or mobile robot).In recent years,visual odometry has been widely used in the field of virtual reality,unmanned driving,mobile robot and other emerging technologies.The traditional visual odometry usually involves a classic ”four-step” approach: 1.First,a specific algorithm is used to calculate the features of an artificial design,which is often called feature extraction;2.Then,the algorithm is used for feature matching between key frames;3.Track the matched features to calculate the pose changes;4.Finally,some optimization strategies are used to reduce the error.The traditional method based on this idea has some problems such as over-dependence on camera parameters and poor adaptability to multiple scenes.In recent years,the vigorous development of deep learning has brought a new way of thinking to the development of VO technology.Compared with the artificial features in the traditional methods,the features extracted by the convolutional neural network are often more reasonable.In addition,the visual odometry based on deep learning method is completely independent of camera parameters,and there is no scale problem in traditional methods.Based on the above advantages,deep learning method is becoming the main research direction of visual odometry.In view of the problems existing in the above traditional methods,this paper has conducted in-depth research on some advanced methods based on deep learning and completed the following works:1.Following the idea of combining deep learning with optical flow method,I conceived the overall algorithm framework.2.The dense optical flow rich in inter-frame motion information is directly input into the convolutional neural network to realize feature extraction,and the full-connection layer is used for pose prediction.Finally,by reducing the function value of the error function,a visual odometry model based on optical flow method and deep learning is trained.3.The influence of two important links in the overall algorithm framework,optical flow extraction and feature extraction,on the performance of the algorithm is analyzed.4.Attention mechanism is introduced into the solution of VO problem,and good results are achieved.In addition,the use of multi-scale pooling technology greatly reduces the dimension of features,thus effectively reducing the number of parameters in the system.To sum up,a visual odometry with both accuracy and speed is designed and implemented gradually in this paper by combining optical flow method and deep learning theory,and the simulation is carried out on the open data set.The rotation error of the model is 0.015 radians per meter,and the translation error percentage is 4.75%.The average time consumption of the algorithm is about 13 milliseconds per frame.