| With the rapid development of UAV technology,more and more UAVs have appeared in the fields of smart logistics,post-disaster rescue,security monitoring,etc.,which puts forward higher requirements for their positioning accuracy.At present,most industrial drones is used GPS technology for navigation,but in some complex environments,GPS information is unreliable or even impossible to obtain,resulting in the failure of drone positioning.Therefore,SLAM system with improved accuracy and robustness has become a hot topic in uav field.For visual SLAM systems,although the theoretical framework is well established in most aspects,autonomous perception in large-scale environments and dynamic scenes with significant appearance changes is still an urgent problem to be solved.This paper studies the SLAM problem in dynamic scenarios.The specific work is as follows:(1)A deep feature-based Super-Net convolutional neural network is proposed to extract the deep features of images.Use a lightweight convolution layer such as dilated convolution to replace ordinary convolution,and optimize the network structure parameters,so that the network can provide a larger receptive field with the same computational cost;shared coding and hierarchical computing are used to extract images in the image.The key points,local feature descriptors and global descriptors are given at the same time,so that the integrated SLAM system can enhance the semantic information processing ability of the image,and the absolute error of pose estimation becomes smaller in the case of changes in the environment and perspective.(2)Using the idea of layers,a relocation method based on global retrieval and group matching is proposed.First,the local features extracted by the convolutional layer are input to the Net VLAD layer to obtain the global feature descriptor describing the feature distribution,and the contour and shape information of the key frame are extracted.For dynamic target objects,the depth information of image features can be roughly retrieved;The frame feature points are matched with all the feature points in the group,a global retrieval is performed to obtain the position hypothesis,and the local features within the position of the candidate frame are retrieved for group matching and classification according to the common view relationship,acting as a fine retrieval of the candidate frame.This coarse-to-fine hierarchical positioning paradigm has better robustness to changes in environment and perspective,can effectively distinguish the semantic information of images,and reduce the workload of back-end optimization.(3)A loop closure detection method based on global and local features is proposed.The DBo W3 bag of words model is added,the global descriptor and local clustering features are added to the LCD,and geometric constraints are added to optimize the online pose estimation.Matching with the pre-trained local feature vocabulary to detect loop closures,a highly reliable closed loop detection method is established,which solves the problem that the DBo W2 bag of words model is prone to initialization failure due to ignoring the spatial relationship of feature points.(4)Selecting the quadrotor UAV as the research object,a set of indoor autonomous inspection UAV system is designed.The SN-SLAM algorithm proposed in this paper is used to locate the UAV,and the Gazebo simulation platform is used to simulate the indoor complex environment information.It is guided by the detection of the target object,adds dynamic obstacles,and uses the path planning and obstacle avoidance algorithms to make it possible in the design.It can make reasonable decisions in a certain obstacle environment,and then efficiently and quickly explore the optimal path between the starting point and the target point,search for the indoor target information and return it to the ground station in real time.After experiments,the algorithm proposed in this paper has better robustness than the ORB-SLAM2 algorithm in the indoor real simulation environment. |
PDF Full Text Request