Research Of Visual Indoor Positioning Method Based On AlexNet Model

In recent years,with the continuous innovation and advancement of information technology and industrial manufacturing capabilities,the application of indoor positioning technology in production and life is increasing.At the same time,it makes indoor positioning technology gain widespread attention that people's demand for location-based services increase gradually.Due to factors such as the walls obstruction in the indoor environment,the robust outdoor positioning method is difficult to apply to indoor positioning scenarios.The scene vision indoor positioning method has become a major research direction in indoor positioning technology because of its simple deployment and wide application scenarios,which can provide semantic information and high speed.This paper expounds the research background and implication of vision indoor positioning technology and related current research situation at home and abroad firstly.The principle and central technology of vision indoor positioning are introduced.It explains the principle and implementation method of scene vision indoor positioning.The research improvements in this paper are mainly in the following aspects: compares several classical corner detection algorithms and proposes an improved image processing method based on feature detection to eliminate random errors;it integrates the specific positioning environment to improve the structure and parameters of the network model and positioning accuracy and efficiency;it considers the complex positioning environment in practical tasks to propose a data augmentation method which can improve the anti-jamming performance of the model.The experiment is carried out on the constructed School dataset,and tests and verifies the effectiveness of the improved method.The scene vision indoor positioning studied in this paper uses the AlexNet convolutional neural network model,the experimental scene selects the campus indoor scene,and uses a manually constructed School dataset which is composed of five common campus indoor scenarios for experimental validation.The research uses the image processing method based on Harris feature detection instead of the original random cropping method,and selects the reserved area based on the number of extracted Harris corner features;For the problem of low accuracy of small-resolution School dataset positioning,the convolution kernel size is adjusted,the effectiveness of the extracted features is improved,the pooling layer and the Dropout layer are increased,and the over-fitting phenomenon is improved;because the visual positioning is easily interfered by factors such as illumination diversification,personnel mobile and sensor rotation,the study simulates the variable factors by means of decentralization,filling blur,rotation,etc.,and improves the data augmentation method of the original model,so that the model can be applied to the complex positioning environment.The experimental results prove that this paper's improved scheme removes the relatively unrelated regions in effect,significantly improves the accuracy of the location recognition of the model.It guarantees good real-time performance and the ability to provide semantic information,reasonably fits the error changes in the visual indoor positioning,improves the anti-interference ability of the positioning system,and has high efficiency and stability.