Application Research In V-SLAM Loop Clusure Detection System With Deep Model Compression

Visual-based Simultaneous Localization and Mapping(VSLAM)refers to the mobile robot uses its own visual sensors to collect environmental image information,then estimate its own movement traj ectory meanwhile obtain a three-dimensional point cloud map of the environment,in other word,iterative realization of localization and mapping.The existing RGBD-SLAM system is based on traditional artificial features for mobile robot trajectory estimation and three-dimensional environment point cloud map construction,which has the characteristics of small calculation,fast and efficient.The RGBD-SLAM system adds a loop closure detection module.The loop closure detection module can effectively eliminate the cumulative error of the sensor,thereby improving the location accuracy of the system.The RGBD-SLAM system uses the smallest spanning tree algorithm for loop closure detection,but because the loop closure detection algorithm still uses traditional artificial features,the detection accuracy of the algorithm is not high when the environment changes complicatedly,and the robustness when responding to environmental changes is poor.The high-precision and fast loop closure detection is of great significance for SLAM location accuracy.Based on this,this paper mainly studies the loop closure detection module of VSLAM,and aims to improve the speed of loop closure detection and the robustness of the algorithm when the environment changes drastically.In loop closure detection,this method has advantages over traditional features due to the use of deep networks for feature learning.on this basis,this paper fuses a loop closure detection algorithm based on deep learning and model compression with the VSLAM system,and proposes a deep-based The model compression network implements a fast and robust VSLAM algorithm for loop closure detection.main tasks as follows.(1)A deep compression loop closure detection algorithm is proposed.The algorithm uses parameter pruning in model compression to accelerate the convolutional neural network,and uses the compressed convolutional neural network to perform feature learning to achieve loop closure detection.The innovation of the algorithm is to apply the pruning of parameters in the model compression to the deep network loop closure detection,and to increase the detection speed of the algorithm by four times without affecting the network loop closure detection performance.(2)Improved VSALM system for deep compression loop closure detection.By combining the proposed deep compression loop closure detection method with RGBD-SLAM system,a VSLAM system based on deep compression loop closure detection is obtained.Compared with the traditional VSLAM system,the improved system has the advantage of using deep networks to learn image features,so that the loop closure detection module can greatly eliminate the cumulative error of the sensor,thereby improving the location accuracy of the VSLAM system.At the same time,the parameter pruning in the model compression is introduced to increase the system speed by ten times on the premise of improving the positioning accuracy of the system.(3)The performance of the proposed deep-compression loop closure detection algorithm and the improved deep-compression loop closure detection VSLAM system performance were experimentally evaluated under standard data sets.The experimental results show that the application of model compression in deep network loop closure detection will not deteriorate the network detection performance,and at the same time greatly improve the algorithm detection speed,the improved deep compression loop closure detection VSLAM system can effectively complete the location and mapping task,the operating speed of the system has also been greatly improved.