Research On Dynamic Target Detection And Removal Method In Unmanned System Environment Map

With the rapid development of related technologies,unmanned systems have been widely used in environment perception,intelligent control,and autonomous navigation.However,the accuracy of the environmental map is crucial for the operation of unmanned systems,as it is used to achieve autonomous navigation and path planning.The presence of dynamic objects in the environment,such as pedestrians and vehicles,can affect the navigation and planning functions of unmanned systems.Therefore,it is necessary to detect and remove dynamic objects in the environmental map to improve the autonomous navigation and path planning capabilities of unmanned systems,and to achieve more efficient,safe and reliable unmanned system applications.The main research content of this thesis is as follows:(1)Explain the perception and mapping system and preprocess the sensor data.Briefly explain the software and hardware environment of the environment perception and mapping system,and preprocess various sensor data,including point cloud filtering and image denoising processing,for subsequent algorithms and data fusion.(2)Research on dynamic target recognition algorithm based on deep learning.Use an instance segmentation algorithm to segment potential dynamic objects on the road,such as cars and pedestrians,and further optimize the detection and segmentation accuracy from the backbone network to the structural parameters based on the traditional Mask R-CNN algorithm.At the same time,based on the SORT target tracking algorithm,add color histogram features for improvement,so that the tracking algorithm can more accurately identify the real dynamic objects on the road.(3)Research on projection-based point cloud and image data association method.Design a time and space synchronization method for lidar and camera.Use the 3D-2D projection method to realize the correlation between point cloud and image data.(4)Research on multi-sensor fusion SLAM system in dynamic scene.A graph based laser SLAM algorithm is adopted,with the front end implementing laser odometry based on the scanning characteristics of the laser radar,and the back end adding IMU pre-integration constraints to improve the accuracy of localization and mapping.Research on target extraction methods based on point cloud geometric features.Investigate a fusion method for point cloud geometric information based on object centroids and instance segmentation information,and remove identified dynamic objects during the mapping process to ensure local mapping consistency.This project mainly studies the problem of map trailing caused by dynamic objects such as pedestrians and vehicles in dynamic scenes.It mainly uses sensor fusion perception of point clouds and images to obtain the point cloud data of dynamic objects in the scene by fusing visual instance segmentation information with lidar clustering information,so as to remove them during mapping,thereby making the point cloud map more accurately reflect the static environment,which is of great significance to improving the accuracy of unmanned system positioning and navigation.