Indoor Navigation Algorithm Of 3D Visual Search And Its System Implementation

Autonomous navigation is the core technology of intelligent mobile robot.To imitate the human cognitive way of visual autonomous navigation is research trend of autonomous navigation.The visual navigation algorithm needs powerful computing to realize the real-time processing of the image,and it is fail to locate itself in the scene where the visual feature is not rich enough.At the same time,the precision of 3D model reconstructed is low due to the large positioning error.In order to solve these problems,this paper proposes a low-cost robot vision autonomous navigation algorithm based on cloud framework,and focuses on the two aspects of high-precision localization and high-precision dense point model.The main contributions are as follows.Aiming at the problem that the autonomous navigation of robot vision is required to carry high cost processor,this paper proposes a low-cost robot vision autonomous navigation algorithm under cloud framework.A three-layer cloud framework is used to construct a system based on visual features and 3D geometric features,as well as dense point model reconstruction,including robot clients,local servers,and private clouds.The high precision localization algorithm and the 3D model reconstruction algorithm are placed on the server with strong computing power.The coarse localization algorithm is placed on the low cost robot client.The algorithm structure of the cloud frame reduces the dependence of the robot on the processor configuration under the premise of ensuring the high precision positioning and modeling.The robot can carry the processor with less memory and weaker computing power to locate itself in real time.This article designs a localization algorithm based on combining visual features and geometric features to solve localization failure when robots go through sense without abundant texture.Robot locates itself based on visual features in a place with rich texture.Otherwise,to estimate the current robot position based on geometric structure.At the same time,we construct a co-visibility graph that represents constrain among key frames,key frames and 3D features to optimize the local key frames and 3D features.If it detects a loop frame,we optimize whole graph.The autonomous navigation system under the cloud framework requires network communication to contact the robot client and the local server.However,it is not avoid that network occurs delay of communication.Ours system have high performance when network is exceptional because of local co-visibility graph in robotic client.In view of the problems of traditional 3D model reconstruction algorithm,this paper put forward the 3D dense model reconstruction algorithm based on key frames.The key frames with high precision are put into the dense model through point fusion.At the same time,the introduction of the depth camera depth measurement error model to estimate confidence of key frames and the closed loop correction improve the precision of dense point model.Due to the direct use of point fusion rather than fusion model based on the TSDF,it reduces the GPU memory consumption and improves the real-time performance of the whole system and obtains the high precision of 3D dense model.We design a system of intelligent robot navigation based on the algorithm.