Research On Key Technologies Of SLAM For Environmental Understanding And Target Tracking

Simultaneous Localization and Mapping(SLAM)have become a hot research topic in robotics.It is the basic module of many emerging technologies,such as autonomous vehicle,robot navigation,unmanned aerial vehicles,virtual reality and augmented reality.However,the environment in which unmanned systems operate is completely unknown in many cases.At this point,unmanned systems require high-precision and highly robust odometers to ensure that they can operate in a long time even when GPS signal are lost.At the same time,the three-dimensional map of the workspace is also very important to robots to achieve positioning,obstacle avoidance,navigation,and autonomous tasks.The content of this paper is how SLAM can better understand the surrounding environment information on the unmanned system,mainly including semantic information,identification of obstacles(positive and negative obstacles),dynamic object recognition,construction of dynamic and static three-dimensional maps,and target tracking.This not only improves the accuracy of SLAM estimation,but also enables unmanned systems to better serve the corresponding work,thereby improving the human-machine interaction ability of unmanned systems and better serving humanity.This article focuses on the key technologies of visual SLAM for research.Visual SLAM has made a series of breakthroughs in its development over the past decade,but its robustness in practical applications is still insufficient.One of the main reasons is the low-level feature extraction,which cannot accurately extract and match feature points in dynamic environments.However,with the significant progress of semantic segmentation technology in computer vision and other related fields,the automatic learning and extraction of relevant features from big data enable visual SLAM to classify most objects through the network.For some issues in visual SLAM systems,semantic segmentation technology can be introduced to enhance or replace performance,thereby improving the performance of visual SLAM systems.At the same time,the detection and tracking of positive and negative obstacles to complex environments can also optimize the robustness of SLAM.Therefore,this article has conducted research on SLAM initialization,negative obstacle detection and tracking,dynamic obstacles in visual SLAM,and the construction of dense point clouds in dynamic environments,and has achieved certain results.The specific content mainly includes:1 To address the issue of better initialization of tightly coupled visual inertia SLAM,this article uses data from four consecutive keyframes for initialization and calculates the biases of accelerometers and gyroscopes,as well as the scale and gravity directions.Based on this,the velocity of the keyframes is obtained.The proposed method was experimentally applied in the Eu Roc database and actual scenarios,and its accuracy was compared with the existing VISLAM system.The experimental results confirm that the developed system has satisfactory accuracy and efficiency.2 In response to address the current issue of insufficient accuracy and real-time detection of negative obstacles,as well as improving SLAM environmental understanding ability.This article extracts 3D negative obstacle candidates through height analysis of point clouds,and obtains 2D negative obstacle candidate regions from3 D negative obstacle candidate regions.At the same time,this article proposes a method similar to RANSAC to solve the noise caused by point cloud matching errors in 3D reconstruction.The proposed method was tested on a 20 Hz camera and the Jetson Xavier NX platform,and image feature tracking were used to ensure detection accuracy and real-time operation.3 In response to the current issue of poor robustness of mainstream SLAM in dynamic scenarios.This article introduces object detection technology into SLAM to achieve better scene perception ability,thus forming a more robust SLAM system.However,object detection relies on deep learning networks that do not perform visual geometric correlation and cannot accurately recognize the motion of objects.Therefore,this article combining object detection technology proposes a pose optimization algorithm based on image segmentation for better long-distance feature matching.A method for selecting the optimal target tracking after identifying moving targets has been proposed.The experiment was validated on 5 sequences on the TUM dataset.Firstly,we conducted ablation experiments on tracking performance.Afterwards,a comprehensive comparison was conducted with ORB-SLAM2,original semantic ORB-SLAM2,RS-SLAM,and DS-SLAM systems,confirming that our algorithm has better accuracy and robustness in dynamic environments.4 To address the issue of inaccurate construction of dynamic objects of dynamic environments in real-time.Using Mask RCNN as semantic prior information and a RGB-D camera model.The spatial coordinate corresponding to the pixels of image are uniformly mapped to a three-dimensional spatial coordinate system to obtain the corresponding dense semantic point cloud.The ground recognition part of the negative obstacles mentioned above is used to extract ground.Using the dynamic object recognition method mentioned above to construct a real-time point cloud of a dynamic environment.Experiments were carried out on the TUM dataset to extract dense point clouds,semantic point clouds,static scene point clouds,dynamic scene point clouds,and 3D frame of dynamic objects.These experiments proved that the proposed system can build clear 3D point cloud images of dynamic environments.Another experiment was conducted that compared with several advanced SLAM systems in order to demonstrate the better performance of our 3D reconstruction based on dynamic tracking optimization.Finally,the running time of each module is displayed.