Point Cloud Map Fusion Algorithm For Multi-UAVs Visual SLAM In Dynamic Scene

In recent years,with the increasing maturity of unmanned aerial vehicle(UAV)technology,the price is gradually low,the performance of the UAV is continuously improved,plus its strong flexibility.Due to the limited time and space constraints,it has been obtained in many fields.One of the basic tasks of the UAV agent is to determine its location information through the sensors carried by the UAV in a given map environment,and the map information is also unknown.Therefore,it is also necessary to construct a 3D map of the environment,which is the simultaneous localization and mapping(SLAM)problem.It is an application of real-time structure from motion(Sf M).When the application scenario is large or large-scale complex,the single UAV cannot realize the positioning and map construction stably.Aiming at the above problems,we mainly study the multi-UAV platform visual SLAM technology based on RGB-D camera sensor,map fusion algorithm and the processing problem of moving targets in dynamic scenes.The main work of this paper is as follows:We study the problems of low efficiency,small amount of tasks and poor system robustness of a single UAV,and construct a system framework of multi-UAV cooperation.Different UAVs are equipped with ASUS Xtion2 depth camera sensor and ORB-SLAM2 algorithm platform to complete their own positioning problems in unknown environment and sub-map construction.On this basis,we propose a multi-point cloud map fusion algorithm based on ICPI(ICP-Improved).Based on the sub-map constructed by each UAV,the matching sub-map coordinates are directly obtained by matching the overlapping regions.The transformation relationship is transformed into the same coordinate system by the rotation matrix and the translation matrix.The improved KD-tree algorithm is used to search the nearest point matching point pairs,and the relationship between the point cloud normal vectors is used to eliminate the false matching points,and the convergence rate of algorithm is improved.We use the ICPI algorithm to perform iterative optimization error,and finally obtain the global 3D point cloud map after fusion.In the process of studying the above SLAM problem,it is found that most mature existing SLAM algorithms assume that the application scenario is static,that is,no moving objects can occur.This strong static assumption limits the application of most visual SLAM algorithms.In reality,real application scenarios such as smart driving and service robots will inevitably have moving objects.Therefore,aiming at the limitations of traditional SLAM application scenarios,we study the processing method of moving objects in SLAM algorithm under dynamic scenarios,and carries out detection and recognition of moving objects based on deep learning R-CNN(Region-CNN)network.The key frame information complements the background after the object is removed,and creates a static map of the moving object in the dynamic scene.Finally,the experimental simulations show that,the proposed method can solve the problem of positioning and mapping of drones in dynamic scenes,which can eliminate moving objects with higher recognition success rate and realize the construction of static maps.The compositional accuracy and trajectory estimation of traditional SLAM are greatly improved.