Research And Application Of Visual Navigation And Target Detection Based On Convolutional Neural Network

With the development of the robot industry,robot-based autonomous navigation and target detection technology has gradually become a hot spot.However,the high computational cost of traditional methods,the complex and varied external environment and the limited computing resources of embedded systems make it a difficult problem.How to realize accurate and efficient autonomous navigation algorithm and detect specific targets at the same time has important research significance for realizing robot autonomous navigation and target detection.In recent years,many scholars have begun to pay attention to the visual navigation algorithm based on convolutional neural network.This method can directly obtain the navigation direction from the original image,and the computational resource requirements are small,and the logic setting problem of the navigation instruction obtained by the environment map is omitted.This paper deeply analyzes the key issues of visual navigation and target detection technology,and carries out research to improve the accuracy and efficiency of the algorithm.This paper proposes a visual navigation method that combines edge features.In order to explore the reason that the visual navigation method based on convolutional neural network can use a single image for navigation,the convolution visualization method is used to interpret and analyze the algorithm.It is found that the algorithm mainly focuses on the edge features of the image.Therefore,this paper proposes to use the canny operator to extract the edge of the image firstly.The extracted edge feature is used as the input of the convolutional neural network together with the original image.This paper improves the network structure to promote the accuracy of the algorithm.Finally,a comparative experiment on the public data set verifies the performance of the algorithm.This paper proposes a multi-task convolutional neural network that realizes both visual navigation and target detection for robot autonomous navigation and target detection.The network structure is that after the input image is extracted by the backbone network,the visual navigation and the target detection branch network are connected to realize visual navigation and target detection respectively.The two branches share the backbone network,improving the efficiency of the algorithm and reducing power consumption,and facilitating the use of the algorithm on the embedded system.In addition,the network detection structure of the target detection algorithm is improved,and the cross-layer connection is used to combine the feature information of different scales for target detection,which improves the algorithm effect.In order to carry out the actual experiment,a data set production method was proposed and a robot data set was made.Experiments on public data sets and self-made data sets show that the proposed method has good accuracy and efficiency.This paper builds a visual navigation and target detection experimental platform for experimental verification of the algorithm.Firstly,based on the ROS robot operating system,the topic nodes and control logic are designed,and the experimental scene is designed for experiment.Then,the experimental platform is equipped with the target detection algorithm proposed in this paper.Experiments are carried out in indoor scenes.Under different angles and different distances,the algorithm can accurately and quickly detect the target,which verifies the accuracy and efficiency performance of the proposed algorithm.Finally,the experimental platform is equipped with the visual navigation and target detection multi-task network proposed in this paper.The experiment is carried out in the indoor corridor scene with obstacles.The experimental platform can correctly visually navigate and operate,and at the same time can detect the target object,which verifies the utility of multitasking networks presented in this paper.