Research On Monocular Visual Odometry And Loop Closure Detection Algorithm Based On Deep Learning

With the development of personal mobile devices in recent years,VR/AR has become a popular research and application area.The key technology to realize AR scene perception and map construction is SLAM.Because personal devices are developing in the direction of miniaturization and portability,visual SLAM has become an important research direction.The two main modules in visual SLAM are visual odometry and loop closure detection,one for estimating interframe pose and the other for eliminating cumulative errors.The traditional SLAM uses manual features to finish the visual odometry and loop closure detection tasks,using feature points and feature descriptors to describe image features.Visual odometry is implemented by pairwise polar geometry,and loop closure is detected by matching the feature descriptors.However,this method relies on camera internal reference and the stability of the scene.The performance of manual features in scenes with large changes in illumination and angle will also be affected to some extent.As a popular field of current research,deep learning can also be applied to visual SLAM with its excellent feature extraction capability in the image field.It can make full use of all the information in the image through end-to-end deep neural networks with a better result in generalization and stability.In this paper,we have done the following work after an in-depth study of deep learning and visual SLAM.(1)For the visual odometry task,proposing a recurrent convolutional neural network CNN-LSTM,which achieves interframe feature extraction of adjacent images by convolutional neural networks,and then performs interframe feature learning between long sequences by recurrent neural networks to achieve multi-frame constraints.The model can directly output the six-degree-of-freedom predicted poses of the camera through end-to-end learning.Compared with the traditional monocular method,the deep learning method has better generalization performance and better performance.(2)Proposing a CNN-CBAM-LSTM network model to optimize the feature extraction capability of CNN by adding CBAM attention module on the basis of recurrent convolutional neural network,which has better pose prediction effect compared with the CNN-LSTM before optimization.(3)For the loop closure detection task,proposing the attention-based VGG convolutional neural network to extract image features instead of traditional feature descriptors,which has a better feature extraction effect in scenes with large changes in lighting,weather,and viewpoint.Using locally sensitive hash method to accelerate feature matching,and designing a hash function based on cosine similarity to detect feature similarity,which is better than the traditional Compared with the traditional double-loop traversal matching.This method has a faster matching speed with losing little accuracy.