WIFI indoor positioning system based on space-time attention convolutional neural network
By using the spatiotemporal attention convolution neural network to process the signal intensity sequence in the WIFI indoor positioning system, the positioning model is formed, and the problems of poor indoor positioning accuracy and large calculation amount in the prior art are solved, and fast and accurate indoor positioning is achieved.
Patent Information
- Application Number
- CN202411575406.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-06
- Publication Date
- 2025-02-28
AI Technical Summary
The existing WIFI indoor positioning methods have problems such as blind spots in indoor environments, slow responses, and poor positioning accuracy. The calculation amount based on fingerprint matching is large, the memory occupies a lot, and the positioning time is long.
The WIFI indoor positioning system based on spatiotemporal attention convolution neural network is adopted to process signal strength sequences through one-dimensional convolutional neural network and attention module to form a positioning model to achieve fast and accurate indoor positioning.
It significantly improves the accuracy and efficiency of indoor positioning, reduces the calculation amount and positioning time, and is suitable for environments with different building structures.
Smart Images
Figure CN119535350A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of indoor positioning technology, and more specifically, the present invention relates to a WIFI indoor positioning system based on spatiotemporal attention convolutional neural network. Background Art
[0002] In recent years, with the rapid development of wireless technology, the Internet of Things, and intelligence, people have an increasingly strong demand for accurate positioning information during indoor activities. At the same time, in some special occasions, such as disaster management, surveillance, object tracking in warehouses, and targeted advertising based on precise locations, the demand is also increasing.
[0003] Although global positioning technology (GPS) is relatively mature and can meet the needs of most people's outdoor activities, in indoor environments, there are many buildings, decorations, and pedestrians, which cause satellite positioning technology to have blind spots in indoor signals, slow response, and poor positioning accuracy.
[0004] At present, the machine learning indoor positioning algorithms based on WIFI fingerprints mainly include KNN algorithm, SVM algorithm, random forest algorithm, and decision tree algorithm. However, these methods need to build a large number of fingerprints first, put them into the fingerprint library, and then match the collected fingerprints with the fingerprints in the fingerprint library to find the fingerprint with the highest matching degree. In order to ensure the positioning accuracy, a large number of fingerprints are required in the fingerprint library, which will take up a lot of memory. In the in-place stage, it is necessary to match with a large number of fingerprints in the fingerprint library, which has the problem of large amount of calculation and long positioning time. Summary of the invention
[0005] The present invention provides a WIFI indoor positioning system based on spatiotemporal attention convolutional neural network, aiming to improve the above problems.
[0006] The present invention is implemented as follows: a WIFI indoor positioning system based on spatiotemporal attention convolutional neural network, the system comprising:
[0007] Multiple WIFI transmitters are set up at different locations indoors, and each WIFI transmitter transmits WIFI signals to the outside in real time;
[0008] The information collection device is fixed on the target object. The signal collection device collects the signal strength Rss of the WIFI signal transmitted by each WIFI transmitting end at the current position to form a signal strength sequence. The signal strength sequence is input into the positioning model, and the positioning model outputs the indoor position of the collection device.
[0009] Furthermore, the method for obtaining the positioning model is as follows:
[0010] Build a one-dimensional convolutional neural network, perform offline training on the one-dimensional convolutional neural network based on training samples, and after the training is completed, test the one-dimensional convolutional neural network based on test samples. After the prediction accuracy of the one-dimensional convolutional neural network reaches the set accuracy requirement, a positioning model is formed;
[0011] The signal strength sequence is used as the feature of the sample, and the corresponding indoor location is used as the label of the sample.
[0012] Furthermore, the one-dimensional convolutional neural network structure includes:
[0013] CNN layer; a channel attention module connected to the CNN layer, and a spatial attention module connected to the channel attention module; the first fully connected layer and the second fully connected layer are connected to the spatial attention module through the first dropout layer and the second dropout layer respectively; the Output layer connected to the first fully connected layer and the second fully connected layer.
[0014] Furthermore, the channel attention module includes:
[0015] A global average pooling layer, a global maximum pooling layer and a multiplier connected to the input end; a two-layer MLP neural network connected to the global average pooling layer and the global maximum pooling layer; an adder connected to the two-layer MLP neural network; a sigmoid activation function connected to the adder, and a multiplier connected to the sigmoid function.
[0016] Furthermore, the width and height of the input feature map F are subjected to global maximum pooling and local average pooling through the global maximum pooling layer and the global average pooling layer to obtain the feature map f m , feature map f a , the feature map f m , feature map f a Input the mlp neural network in sequence to get the feature map f m ′ and feature map f a ′, the feature map f m ′ and feature map f a ′After the input adder is added, the sigmoid activation operation is performed to generate the feature map f c , feature map f c Input multiplier, feature map f c Multiply it with the feature map F to get the feature map F_c containing channel attention.
[0017] Furthermore, the spatial attention module includes:
[0018] The global average pooling layer, the global maximum pooling layer and the multiplier connected to the input end; the Concatenation layer connected to the global average pooling and the global maximum pooling; the convolution layer Conv connected to the Concatenation layer; the sigmoid activation function connected to the convolution layer, and the multiplier connected to the sigmoid activation function.
[0019] Furthermore, after the global average pooling layer and the global maximum pooling layer perform global average pooling and global maximum pooling on the width and height of the input feature map F_c, the feature map f d , feature map f e , the feature map f d , feature map f e Input Concatenation layer, feature map f d , feature map f e Perform stacking operations to form feature maps f de , feature map f de Input the convolution layer and obtain the feature map f through a one-dimensional convolution operation. k , feature map f k After sigmoid activation operation, the feature map f is generated s , the feature map f s Input multiplier, feature map f s Multiply it with the feature map F_c to get the feature map F_s of the channel space attention.
[0020] Furthermore, the system also includes:
[0021] The database is used to store the constructed samples and divide the samples into training samples and test samples according to the set ratio.
[0022] Furthermore, the sample construction process is as follows:
[0023] After arranging the WIFI transmitter in the indoor space, the signal strength Rss of the WIFI signal transmitted by each WIFI transmitter and the device identification of the WIFI transmitter are collected at different locations in the indoor space. All signal strengths Rss are organized into a signal strength sequence according to the set device identification order, and the indoor position corresponding to the signal strength sequence is marked. The indoor position is composed of the building number and floor number to complete the sample construction.
[0024] The positioning model of the present invention adopts the strategy of offline training and online use, and during use, it only needs to input the currently collected signal strength sequence into the trained positioning model to obtain the corresponding positioning information, which greatly improves the problems of large calculation amount and long positioning time in the existing positioning method;
[0025] In addition, the two-dimensional spatiotemporal attention module is improved to form a one-dimensional spatiotemporal attention module that can process one-dimensional signal intensity sequences. The one-dimensional spatiotemporal attention module focuses on features with larger signal intensity in significant signal intensity sequences, and the one-dimensional channel attention module focuses on better feature maps in multiple feature maps. This makes the positioning method proposed in the present invention better applicable to different building structures in the environment. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 A schematic diagram of the structure of a WIFI indoor positioning system based on a spatiotemporal attention convolutional neural network provided by an embodiment of the present invention;
[0027] Figure 2 A schematic diagram of the structure of a one-dimensional convolutional neural network provided in an embodiment of the present invention. Implementation
[0029] The specific implementation modes of the present invention are further explained in detail below by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.
[0030] Figure 1 A schematic diagram of the structure of a WIFI indoor positioning system based on a spatiotemporal attention convolutional neural network provided in an embodiment of the present invention. For ease of explanation, only the part related to the embodiment of the present invention is shown. The system includes:
[0031] Multiple WIFI transmitters are set up at different locations indoors, and each WIFI transmitter transmits WIFI signals to the outside in real time;
[0032] The information collection device is fixed on the target object. The signal collection device collects the signal strength Rss of the WIFI signal transmitted by each WIFI transmitting end at the current position to form a signal strength sequence. The signal strength sequence is input into the positioning model, and the positioning model outputs the indoor position of the collection device.
[0033] In an embodiment of the present invention, the positioning model is constructed based on a one-dimensional convolutional neural network. After the one-dimensional convolutional neural network is constructed, samples are constructed, and the constructed samples are placed in a database. The samples in the database are divided into training samples and test samples according to a set ratio. The training samples are used to train the positioning model, and the test samples are used to test the prediction accuracy of the positioning model. The one-dimensional convolutional neural network is trained based on the training samples. After the training is completed, the one-dimensional convolutional neural network is tested based on the test samples. After the prediction accuracy of the one-dimensional convolutional neural network reaches the set accuracy requirement, a positioning model is formed.
[0034] In the embodiment of the present invention, the sample construction process is specifically as follows:
[0035] After arranging the WIFI transmitter in the indoor space to be tested, the signal strength Rss of the WIFI signal transmitted by each WIFI transmitter and the device identification of the WIFI transmitter are collected at different positions in the indoor space. All signal strengths Rss are combined into a signal strength sequence according to the set device identification sequence, and the indoor position corresponding to the signal strength sequence is marked. The indoor position is composed of the building number and the floor number to complete the construction of the sample. The sample uses the signal strength sequence as a feature and the indoor position corresponding to the feature as a label. The constructed sample is put into the database. In order to locate the target object more accurately, the positioning coordinates are used as the label of the sample, that is, the signal strength sequence is used as a feature, and the indoor position and positioning coordinates corresponding to the feature are used as labels to train the constructed one-dimensional convolutional neural network offline.
[0036] It should be noted that at the current location of the target object, the signal acquisition device is not necessarily in the coverage area of the WIFI signals of all WIFI transmitters. There is a problem that the WIFI signals of some WIFI transmitters cannot be received. When constructing the signal strength sequence, for the WIFI transmitter that does not receive the WIFI signal, the signal strength of the corresponding position in the signal strength sequence is set to zero.
[0037] In the embodiment of the present invention, channel attention and spatial attention are simultaneously introduced into the one-dimensional convolutional neural network, which can pay more attention to eye-catching features. The one-dimensional convolutional neural network structure designed by the present invention is as follows: Figure 2 As shown, the structure includes:
[0038] CNN layer; a channel attention module connected to the CNN layer, and a spatial attention module connected to the channel attention module; the first fully connected layer and the second fully connected layer are connected to the spatial attention module through the first dropout layer and the second dropout layer respectively; the Output layer connected to the first fully connected layer and the second fully connected layer.
[0039] In an embodiment of the present invention, the channel attention module includes:
[0040] A global average pooling layer, a global maximum pooling layer and a multiplier connected to the input end; a two-layer MLP neural network connected to the global average pooling layer and the global maximum pooling layer; an adder connected to the two-layer MLP neural network; a sigmoid activation function connected to the adder, and a multiplier connected to the sigmoid function.
[0041] The input feature map F (H×1×C) is subjected to global maximum pooling layer and global average pooling layer to perform global maximum pooling and local average pooling on the width and height of the feature map F, respectively, to obtain two 1×1×C feature maps f m , feature map fa , the feature map f m , feature map f a Input the mlp neural network in sequence to obtain two new 1×1×C feature maps f m ′ and feature map f a ′, the feature map f m ′ and feature map f a ′After the input adder is added, the sigmoid activation operation is performed to generate the feature map f c , feature map f c Input multiplier, feature map f c Multiply it with the feature map F to get the feature map F_c containing channel attention; among them, the number of input nodes and output nodes of the mlp neural network are the same
[0042] In an embodiment of the present invention, the spatial attention module includes:
[0043] The global average pooling layer, the global maximum pooling layer and the multiplier connected to the input end; the Concatenation layer connected to the global average pooling and the global maximum pooling; the convolution layer Conv connected to the Concatenation layer; the sigmoid activation function connected to the convolution layer, and the multiplier connected to the sigmoid activation function.
[0044] Input feature map F_c (H×1×C), after global average pooling layer and global maximum pooling layer respectively perform global average pooling and global maximum pooling on the width and height of feature map F_c, two C×1×1 feature maps f d , feature map f e , the feature map f d , feature map f e Input Concatenation layer, feature map f d , feature map f e Perform a stacking operation to form a C×2×1 feature map f de , feature map f de Input the convolution layer, and obtain the C×1×1 feature map f through a one-dimensional convolution operation. k , feature map f k After sigmoid activation operation, the feature map f is generated s , the feature map f s Input multiplier, feature map f s Multiply it with the feature map F_c to get the feature map F_s of the channel space attention.
[0045] In order to enable the positioning model to correctly distinguish the RSSI strength features at different locations, the CNN layer uses a one-dimensional convolutional layer to extract features from the input signal strength sequence. Since there may be multiple features at different locations, multiple convolutional layers of different lengths are used for extraction, aiming to extract key information that distinguishes different locations. For example, the input data 1×1×520 passes through a 1D-CNN layer with 64 filters (length 13, step size 2) to form a 64×1×254 local feature, using the ReLU activation function.
[0046] After being processed by the above-mentioned channel attention module and spatial attention module, the features of channel attention and spatial attention are obtained. After the features are flattened, they are input into two fully connected layers, one of which is used for the classification of buildings and floors, and the other is used for the regression of the coordinates of the floor of the building. Each fully connected layer is a shallow neural network containing only one hidden layer.
[0047] The present invention uses the publicly available UJIIndoorLoc dataset, which was provided in the Evaluation Environment Assisted Living (EvAAL) competition of the 2015 International Conference on Indoor Positioning and Indoor Navigation (IPIN). The dataset contains 19,937 training data and 1,111 verification data.
[0048] The test environment of the experiment of the present invention is a Lenovo XiaoXin-15ARE 2020 notebook with an AMD Ryzen 7 4800U processor. The Pytorch deep learning framework of Anaconda is used, and the versions are Anaconda 4.9.2 and Pytorch 1.4.0.
[0049] This paper compares the existing building floor prediction task and location prediction task based on the UJIindoorLoc dataset. Taking the SAE+ classifier and the 2D-CNN+ classifier SAE+1D-CNN+ classifier as references, Table 1 shows the positioning accuracy comparison results:
[0050] Table 1 Comparison of positioning results
[0051]
[0052] Based on the above Table 1, it can be seen that the positioning accuracy of the technical solution provided by the present invention is significantly better than that of the SAE+ classifier and the 2D-CNN+ classifier, and is comparable to the positioning accuracy of the SAE+1D-CNN+ classifier, but its average positioning error is significantly lower than that of the SAE+1D-CNN+ classifier.
[0053] The present invention has been described exemplarily. Obviously, the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.
Claims
1. A WIFI indoor positioning system based on spatiotemporal attention convolutional neural network, characterized in that: The system comprises: Multiple WIFI transmitters are set up at different locations indoors, and each WIFI transmitter transmits WIFI signals to the outside in real time; The information collection device is fixed on the target object. The signal collection device collects the signal strength Rss of the WIFI signal transmitted by each WIFI transmitting end at the current position to form a signal strength sequence. The signal strength sequence is input into the positioning model, and the positioning model outputs the indoor position of the collection device.
2. The WIFI indoor positioning system based on spatiotemporal attention convolutional neural network as claimed in claim 1, characterized in that: The method for obtaining the positioning model is as follows: Build a one-dimensional convolutional neural network, perform offline training on the one-dimensional convolutional neural network based on training samples, and after the training is completed, test the one-dimensional convolutional neural network based on test samples. After the prediction accuracy of the one-dimensional convolutional neural network reaches the set accuracy requirement, a positioning model is formed; The signal strength sequence is used as the feature of the sample, and the corresponding indoor location is used as the label of the sample.
3. The WIFI indoor positioning system based on spatiotemporal attention convolutional neural network as claimed in claim 1, characterized in that: The one-dimensional convolutional neural network structure includes: CNN layer; a channel attention module connected to the CNN layer, and a spatial attention module connected to the channel attention module; the first fully connected layer and the second fully connected layer are connected to the spatial attention module through the first dropout layer and the second dropout layer respectively; the Output layer connected to the first fully connected layer and the second fully connected layer.
4. The WIFI indoor positioning system based on spatiotemporal attention convolutional neural network as claimed in claim 3, characterized in that: The channel attention module includes: A global average pooling layer, a global maximum pooling layer and a multiplier connected to the input end; a two-layer MLP neural network connected to the global average pooling layer and the global maximum pooling layer; an adder connected to the two-layer MLP neural network; a sigmoid activation function connected to the adder, and a multiplier connected to the sigmoid function.
5. The WIFI indoor positioning system based on spatiotemporal attention convolutional neural network as claimed in claim 4, characterized in that: After the global maximum pooling layer and the global average pooling layer, the width and height of the input feature map F are pooled globally and locally, and the feature map f is obtained. m , feature map f a , the feature map f m , feature map f a Input the mlp neural network in sequence to get the feature map f m ′ and feature map f a ′, the feature map f m ′ and feature map f a ′After the input adder is added, the sigmoid activation operation is performed to generate the feature map f c , feature map f c Input multiplier, feature map f c Multiply it with the feature map F to get the feature map F_c containing channel attention.
6. The WIFI indoor positioning system based on spatiotemporal attention convolutional neural network as claimed in claim 3, characterized in that: The spatial attention module includes: The global average pooling layer, the global maximum pooling layer and the multiplier connected to the input end; the Concatenation layer connected to the global average pooling and the global maximum pooling; the convolution layer Conv connected to the Concatenation layer; the sigmoid activation function connected to the convolution layer, and the multiplier connected to the sigmoid activation function.
7. The WIFI indoor positioning system based on spatiotemporal attention convolutional neural network as claimed in claim 5, characterized in that: After the global average pooling layer and the global maximum pooling layer perform global average pooling and global maximum pooling on the width and height of the input feature map F_c, the feature map f d , feature map f e , the feature map f d , feature map f e Input Concatenation layer, feature map f d , feature map f e Perform stacking operations to form feature maps f de , feature map f de Input the convolution layer and obtain the feature map f through a one-dimensional convolution operation. k , feature map f k After sigmoid activation operation, the feature map f is generated s , the feature map f s Input multiplier, feature map f s Multiply it with the feature map F_c to get the feature map F_s of the channel space attention.
8. The WIFI indoor positioning system based on spatiotemporal attention convolutional neural network as claimed in claim 1, characterized in that: The system further comprises: The database is used to store the constructed samples and divide the samples into training samples and test samples according to the set ratio.
9. The WIFI indoor positioning system based on spatiotemporal attention convolutional neural network as claimed in claim 1, characterized in that: The sample construction process is as follows: After arranging the WIFI transmitter in the indoor space, the signal strength Rss of the WIFI signal transmitted by each WIFI transmitter and the device identification of the WIFI transmitter are collected at different locations in the indoor space. All signal strengths Rss are organized into a signal strength sequence according to the set device identification order, and the indoor position corresponding to the signal strength sequence is marked. The indoor position is composed of the building number and floor number to complete the sample construction.