Research And Application Of Convolutional Neural Network In SLAM Visual Odometer

Visual odometer is the key technology in simultaneous localization and mapping(SLAM)system.The visual odometer in SLAM includes feature extraction,feature matching and inter frame estimation.The performance of SLAM system is affected by the stability of feature extraction and the accuracy of inter-frame estimation.Due to the influence of uncertain factors such as noise and illumination,and the constraints of feature point type and quantity,the traditional artificial feature extraction method has some limitations.Recently,scholars have begun to study the method of convolutional neural network,which is different from traditional artificial methods to extract features and estimate inter-frame motion.However,the convolutional neural network is complex and requires a large training data set to train.At present,there is no training data set to solve the visual odometer.In this thesis,the training data set of feature extraction and inter frame estimation is designed to train the neural network.The first neural network completed the extraction of feature points.The second neural network matched the extracted feature points and estimated the inter frame transformation relationship.The first neural network processes the single image in the sequence of images and extracts the corner features of the images.The comer location is clear and reliable,which is a typical stable and reliable feature.While there are too many feature points extracted by traditional artificial methods,and it extracts the class eomer feature rather than the accurate comer feature.In this thesis,the classical VGG convolutional neural network is adjusted to extract the feature points of the image.In order to extract stable comer feature,the OpenCV function is used to generate tens of thousands of geometric images to train the network structure.The network is compared with the classic corner detector.And it is found that the performance of feature point extraction is significantly improved when the image with noisyThe second neural network is to process two consecutive pictures in the image sequence and estimate the relationship between frames.After the feature points are extracted,the point pairs need to be matched,and the inter-frame transformation need to be estimated.In this thesis,the structure of neural network is improved on the basis of VGG neural network.The better model is obtained by training the network parameters.In the process of network parameter training,a big obstacle is that there is no complete set of data.Most data sets only contain picture sequences,but lack of real inter frame transformation relationship.In this thesis,some possible motion trajectories of the camera are simulated in a three-dimensional commercial complex.The picture each time the camera observes is output as a neural network training data set.Through experiments,the network model is proved to be better in matching and frame estimation.Finally,in this thesis,questions and lessons learned during the research of the project are summarized.Future convergence of convolutional neural networks in slam is expected.