Research And Terminal Implementation Of Visual SLAM Based On Deep Learning

In recent years,with the rapid development of technologies such as intelligent robots and unmanned vehicles,SLAM systems are playing an increasingly important role as the basis for positioning and navigation of intelligent robots and unmanned ve-hicles.The use of low-cost vision sensors to build visual SLAM systems is also playing an increasingly important role,which then became a hot issue.Nevertheless,the feature points and descriptors obtained by the manual design algorithm have poor robustness,and the accuracy is poor under the conditions of illumination changes,angle changes and low texture.Therefore,the visual SLAM system obtaining feature points and de-scriptors by adopting convolutional neural networks has significant research value.Accordingly,this thesis takes the visual SLAM based on deep learning as the re-search object,and carries out related research aiming at the feature point and descriptor extraction network and model compression.The results of this thesis are as follows:1.Aiming at the feature points and descriptor algorithms in visual SLAM,this thesis proposes a feature point and descriptor extraction network based on depthwise separable convolution.On the basis of the Superpoint network,this thesis improves the upsampling method and loss function form of the Superpoint network descriptor decoder,applies the depthwise separable convolution to the Superpoint network,and finally changes the number of network layers,convolution kernel dimensions and down-sampling methods.The experimental results show that the proposed network is faster in extracting feature points and descriptors,and the extraction accuracy is similar,which proves the effectiveness of the method.2.Aiming at the problem that the feature point and descriptor extraction network runs slowly on devices with limited computing resources,this thesis proposes a model compression algorithm combining network pruning and knowledge distillation.This thesis improves on the channel pruning algorithm to apply it to deep separable con-volution and changes the pruning process so that the pruning algorithm can prune the network of this thesis.Aiming at the problem of excessively high pruning compression rate that leads to a significant decrease in model performance,this thesis uses knowl-edge distillation to improve the accuracy of the network extracting feature points and descriptors after pruning.This thesis uses the model provided in the Superpoint the-sis as the teacher network,and the pruned model as the student network.Finally,the accuracy of the network is only slightly reduced while the compression rate is high.3.This thesis applies the proposed lightweight network to the feature point and descriptor extraction process of ORB-SLAM2 and realize a visual SLAM system based on deep learning.The experimental results show that the accuracy of the SLAM tra-jectory estimation in this thesis is higher than that of the ORB-SLAM2 system,which verifies the effectiveness of the system in this thesis.