Study On The Indoor Positioning Algorithm Based On Improved Convolution Neural Network

With the high-speed development of science and technology in recent years,human beings are not satisfied with the era of industrial machines instead of human simple requirements,but to let the machine have similar to the human brain instead of we handling the problems of the complexity of life,artificial intelligence arises at the historic moment.The traditional navigation system is implemented based on satellite navigation,but was unable to realize the indoor positioning navigation is due to the complexity of the indoor environment more,reach the ground weak signal and cause cannot penetrate the wall.Common indoor navigation and positioning method has wi-fi,bluetooth,infrared,ultra broadband,RFID,ZigBee and ultrasonic positioning,etc.,in this paper,the main method is based on the improvement of the convolution of the neural network monocular visual indoor location,the method is in view of the traditional method of cost is high,the anti-interference ability is poor,weak robustness and bad compatibility problems and put forward the intelligent auxiliary positioning method,the traditional computer vision camera pose estimation method with complex background in the image features are extracted angular point,affected by the angular point of interest is very serious,so this paper joined the convolutional neural network algorithm in complex indoor scene joined the regional limit,complete camera pose estimation in the area of the interested in better realized to implement low cost,high precision and more stable navigation service.The main contents of this subject are the following two aspects:(1)To improve the traditional convolutional neural network,implement the end-to-end model training and test,and complete the identification and classification of indoor single target memory.Traditional target identification method is very difficult to under the complex indoor scene recognition and grab single objective content area of the image,and convolution neural network able to learn from a large number of sample data to the image characteristic of the deep,abstraction,in the field of image recognition and classification showed significant advantages.In this paper,the traditional convolution neural network is improved to realize the integration of classification and detection,mainly including: different scale feature detection,multi-feature fusion,end-to-end integration.In this paper,the detection accuracy of the model is improved,and the detection speed is improved,and the recognition and grasping of the interest area in the image is realized.The test accuracy of the PASCAL VOC dataset was 73.1%,and the accuracy of the data set for this task was 98.3%,and the test speed was 30 frames.(2)In the above work is completed,contour approximation method were used respectively as well as Harris corner detection methods,such as the single objective record angular point grab,determine its position in the image,and then turn the camera pose estimation problem for PNP equations to solve the problem,and the traditional RPnP,EPnP and CEPPnP algorithm were analyzed,and to seek the optimal captured on camera pose estimation algorithm for the current position of the world coordinates,the experimental results show that the proposed localization algorithm has higher stability,single image detection precision error no more than 0.01 m in xyz three directions,video streaming 30 frames error under the condition of not high 0.08 m,the indoor positioning accuracy.