A visible light indoor positioning method based on a multi-branch residual convolutional network

Abstract

The application discloses a visible light indoor positioning method based on a multi-branch residual convolutional network, and relates to the technical field of indoor positioning, in particular to a visible light indoor positioning method based on a multi-branch residual convolutional network. The method comprises the following steps: acquiring a visible light RSS data graph and constructing a training set; training a preset positioning model based on the training set to obtain a trained positioning model; the preset positioning model comprises a residual convolutional network and a multi-branch feature fusion network; acquiring a to-be-measured RSS data graph and inputting the to-be-measured RSS data graph into the trained network model to obtain corresponding position information. By using the application, the positioning influence caused by unknown directions and dynamic changes of the environment is solved, and a robust positioning solution is achieved. The application can be widely applied to the field of visible light indoor positioning.

Description

TECHNICAL FIELD

[0001] The application relates to the field of visible light indoor positioning, and in particular to a visible light indoor positioning method based on a multi-branch residual convolutional network. BACKGROUND

[0002] In recent decades, indoor positioning technology has developed rapidly, and methods for indoor positioning using WiFi positioning and Bluetooth positioning have appeared, but the accuracy is limited. With the rapid development of light-emitting diode technology, visible light-based indoor positioning technology has attracted more and more attention. In particular, with the rapid development of deep learning, the visible light-based indoor positioning method combined with the deep learning mode greatly improves the positioning performance. However, whether the positioning is based on a traditional signal propagation model or a fingerprint-based visible light positioning, the position information of the target is estimated under the condition that the direction information of the target is known. When the positioning performance is interfered by factors such as light shielding and dynamic changes of environmental parameters, the positioning performance will sharply decrease. SUMMARY

[0003] The application aims to provide a visible light indoor positioning method based on a multi-branch residual convolutional network, to solve the positioning influence caused by unknown directions and dynamic changes of the environment, and to achieve a robust positioning solution.

[0004] The first technical solution adopted by the application is a visible light indoor positioning method based on a multi-branch residual convolutional network, comprising the following steps:

[0005] Obtaining a visible light RSS data graph and constructing a training set;

[0006] Training a preset positioning model based on the training set to obtain a trained positioning model;

[0007] The preset positioning model comprises a residual convolutional network and a multi-branch feature fusion network;

[0008] Obtaining a to-be-measured RSS data graph and inputting the RSS data graph into the trained network model to obtain corresponding position information.

[0009] Further, the step of obtaining a visible light RSS data graph and constructing a training set specifically comprises:

[0010] Simulating indoor visible light propagation based on a Lambert model under preset conditions;

[0011] Obtaining an RSS data graph under the conditions of dynamic changes of bulb power, Lambert coefficient and fast decay system;

[0012] Constructing a training set according to the RSS data graph and real position labels.

[0013] Further, the preset conditions are:

[0014] LED lamps are evenly distributed on the roof, and the direction is vertically downward;

[0015] The positioning targets are randomly distributed in the room positions and randomly distributed in the directions;

[0016] The central visible light receiver is vertically upward to the ground, and the remaining visible light receivers are evenly distributed on the edge of the circle with the central receiver as the core, and form a 45-degree angle with the direction of the central receiver.

[0017] Further, the step of training the preset positioning model based on the training set to obtain the trained positioning model specifically includes:

[0018] Training the preset positioning model based on the training set;

[0019] Input the normalized RSS data graph into the residual convolutional network to extract different levels of feature information graphs;

[0020] Respectively send the feature information graphs of different levels into the multi-branch feature fusion network to fuse the feature texture information of the shallow and deep layers to obtain the position estimation information;

[0021] Combine the real position labels in the training set to adjust the parameters of the residual convolutional network and the multi-branch feature fusion network to obtain the trained positioning model.

[0022] Further, the residual convolutional network includes an initial convolutional layer and a residual learning module, and the step of inputting the normalized RSS data graph into the residual convolutional network to extract different levels of feature information graphs specifically includes:

[0023] Input the normalized RSS data graph into the residual convolutional network;

[0024] Extract the primary texture information of the RSS data graph based on the initial convolutional layer;

[0025] The residual learning module based on the shortcut connection learns the residual information between the input features and the output features;

[0026] Extract high-level features based on the residual learning module and transmit the high-level features to the next level of residual learning module.

[0027] Further, the multi-branch feature fusion network includes a multi-branch network, a global average pooling layer, and a full connection layer, and the step of respectively sending the feature information graphs of different levels into the multi-branch feature fusion network to fuse the feature texture information of the shallow and deep layers to obtain the position estimation information specifically includes:

[0028] Input the feature maps mentioned by the plurality of residual convolutional networks into the multi-branch feature fusion network for fusion;

[0029] The feature map is information extraction mapping based on a multi-branch network.

[0030] The feature map is globally averaged by a global average pooling layer to obtain feature point data.

[0031] The feature point data is spliced and stacked based on a fully connected layer, and is mapped with the actual position of the target.

[0032] Further, the acceptance intensity value of each data point of the visible light RSS data graph is as follows:

[0033]

[0034] In the above formula, z m,n represents the acceptance intensity value between the nth visible light receiver and the mth LED, Ψ m,n represents a system parameter, r m represents the Lambert coefficient of the mth LED lamp, x n represents the three-dimensional position of the receiver, p m represents the three-dimensional position of the LED lamp, ф m,n represents the LED emission angle, θ m,n represents the receiver incident light angle, ∈ m,n represents environmental noise.

[0035] The method and system have the following beneficial effects: the method effectively mines and extracts feature texture information of the visible light transmission process through a deep residual network, effectively avoids the performance degradation phenomenon of the deep network, and further, fuses shallow and deep feature information through a multi-branch link to obtain more rich internal feature information, so as to solve the positioning difficulty problem caused by the unknown direction of the object and the dynamic change of the environmental parameters. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a step flow chart of a visible light indoor positioning method based on a multi-branch residual convolutional network according to the present application;

[0037] Figure 2 is a multi-branch residual convolutional network model according to a specific embodiment of the present application;

[0038] Figure 3 is a visible light propagation simulation system built according to a specific embodiment of the present application;

[0039] Figure 4 is a positioning accuracy comparison diagram of the method of the present application and other visible light positioning methods under different signal-to-noise ratios. DETAILED DESCRIPTION

[0040] The application will be further described in detail below in combination with the drawings and specific embodiments. For the step numbers in the following embodiments, they are only set for the convenience of description and explanation, and the order between the steps is not limited in any way, and the execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0041] With reference to Figure 1 and Figure 2 , the application provides a visible light indoor positioning method based on a multi-branch residual convolutional network, which comprises the following steps:

[0042] S1, acquiring a visible light RSS data graph and constructing a training set;

[0043] S2, training a preset positioning model based on the training set to obtain a trained positioning model;

[0044] The preset positioning model comprises a residual convolutional network and a multi-branch feature fusion network;

[0045] S3, acquiring a to-be-tested RSS data graph and inputting it into the trained network model to obtain corresponding position information.

[0046] Further as a preferred embodiment of the method, the step of acquiring a visible light RSS data graph and constructing a training set specifically comprises:

[0047] Under preset conditions, simulating indoor visible light propagation based on a Lambert model;

[0048] Under the conditions of dynamic changes of bulb power, Lambert coefficient and fast decay system, acquiring an RSS data graph;

[0049] Normalizing the RSS data graph and the real position label, and constructing a training set.

[0050] Specifically, according to the Lambert model, the indoor visible light propagation process is simulated, the LED lamps are uniformly distributed on the top of the room, the positions of different LED lamps are optimized and adjusted, and the directions of different visible light receivers are adjusted, so as to obtain an RSS data training graph with rich texture feature information, and the dynamic random environmental parameter influence is increased to make the data closer to the real situation, and the three-dimensional position information of the data graph is recorded as a label for subsequent network training.

[0051] For the received visible light RSS data graph, the size is set to 28x56, the number of LED is 56, the number of visible light receivers is 28, the size of the room is 20x20x5m, and the received intensity value formula of each data point of the visible light RSS data graph is as follows:

[0052]

[0053] In the formula, z m,n represents the receiving intensity value between the nth visible light receiver and the mth LED, Ψ m,n represents the system parameters, mainly affected by the fast decay coefficient g and the LED power W m Dynamic fluctuations, the rest of the receiver aperture, concentrator gain and optical filtering gain and other parameters are set to constant, r m represents the Lambert coefficient of the mth LED lamp, x n represents the three-dimensional position of the receiver, p m represents the three-dimensional position of the LED lamp, ф m,n represents the LED emission angle, θ m,n represents the angle of incident light on the receiver, ∈ m,n represents the ambient noise

[0054] Lambert coefficient r m The value range of is that the mean value is 1.5 and the variance is 0.05; the bulb power is that the mean value is 2.5 and the variance is 0.05; the fast decay coefficient is that the mean value is 0.8 and the variance is 0.05, and the signal-to-noise ratio SNR value range is [-30, 40], to obtain the RSS data graph close to the real environment.

[0055] Further, as a preferred embodiment of the method, the preset condition is:

[0056] The LED lamps are evenly distributed on the roof, and the direction is vertically downward;

[0057] The positioning target is randomly distributed in the room position, and the direction is randomly distributed. The direction of the positioning object is based on the direction of the central receiver.

[0058] The direction of the central visible light receiver is vertically upward, and the remaining visible light receivers are evenly distributed on the edge of the circle with the central receiver as the core, and form a 45-degree angle with the direction of the central receiver.

[0059] Specifically, the visible light system constructed refers to Figure 3 .

[0060] Further, as a preferred embodiment of the method, the step of training the preset positioning model based on the training set to obtain the trained positioning model specifically includes:

[0061] Training the preset positioning model based on the training set;

[0062] Input the RSS data graph after normalization into the residual convolution network to extract different levels of feature information graph;

[0063] The feature information graphs of different levels are respectively sent into the multi-branch feature fusion network to fuse the feature texture information of the shallow layer and the deep layer to obtain the position estimation information.

[0064] The residual convolution network and the multi-branch feature fusion network are adjusted in parameters in combination with the real position labels in the training set to obtain the trained positioning model.

[0065] In the training process, the model parameters associated to different branch networks in the residual network are fused with the gradient information of different branch networks in the gradient update.

[0066] Further as the preferred embodiment of the method, the residual convolution network comprises an initial convolution layer and a residual learning module, and the step of inputting the normalized RSS data graph into the residual convolution network to extract the feature information graphs of different levels specifically comprises:

[0067] The visible light RSS data graph in the training set is input into the residual convolution network.

[0068] The initial convolution layer is used to extract the primary texture information of the RSS data graph.

[0069] Specifically, the initial convolution layer is set to have 16 kernels, and the convolution kernel size is 5x5. Different convolution kernels extract multiple shallow feature information of the input data graph to enrich the original image information.

[0070] The residual learning module is used to extract the advanced features, and the advanced features are transmitted to the joint layer and the next residual learning module.

[0071] Specifically, the residual learning module is used to extract the advanced features, and the residual learning module is composed of two "convolution layer + batch normalization layer" and one rectified linear unit. Each residual convolution block is connected with an identity mapping shortcut connection. The residual learning method is used to replace the direct learning feature mapping method to improve the feature extraction capability.

[0072] The first residual convolution learning module has 16 convolution kernels, and the convolution kernel number of each subsequent module has a multiple relationship with the first module. The more the convolution kernel number is, the more the feature maps are, which is beneficial to the global average pooling processing of the subsequent feature maps.

[0073] The residual learning module is used to extract the advanced features, and the advanced features are transmitted to the next residual learning module.

[0074] Specifically, the initial convolution layer is used to expand the data information, and then multiple residual convolution modules are used to extract and transmit the data feature maps generated by each module. The linear rectification layer is used to enhance the network mapping capability. The identity mapping shortcut connection is added in each residual convolution module to effectively prevent the deep gradient vanishing phenomenon and accelerate the model training speed.

[0075] Further as the preferred embodiment of the method, the multi-branch feature fusion network comprises a multi-branch network, a global average pooling layer and a full connection layer, the step of sending the feature information maps of different levels into the multi-branch feature fusion network respectively, fusing the feature texture information of the shallow layer and the deep layer, and obtaining the position estimation information, specifically comprises:

[0076] The feature maps mentioned in the plurality of residual convolutional networks are input into the multi-branch feature fusion network for fusion;

[0077] Information extraction and mapping of the feature maps are performed based on the multi-branch network;

[0078] Specifically, the input is the output feature map of each residual learning module, and one branch is set for the feature map of each residual learning module for information extraction and mapping

[0079] Global average pooling of the features is performed based on the global average pooling layer to obtain the feature point data;

[0080] Specifically, global average pooling is performed on each input feature map, one feature map is globally pooled into one value, which can effectively prevent overfitting phenomenon and reduce the number of training parameters and accelerate the model training speed.

[0081] The feature point data are spliced and stacked based on the full connection layer, and are mapped with the actual position of the target.

[0082] Specifically, after global pooling, the full connection layer is used to map the feature point data to the target position, a splicing process of the multi-branch is performed in the last layer, the features of each branch are fused, and the three-dimensional position information of the target is obtained, and the feature point data splicing and stacking formula is as follows:

[0083] F=[F1,F2,…,Fi] i ]

[0084] Wherein, F1, F2, …, Fi represent the feature data point sets generated by the 1st, 2nd, …, i-th branch, and F represents the fused and spliced feature point data. i

[0085] The network is composed of a residual convolutional network and a multi-branch network, the two networks are trained in a joint manner, the output layer is shared by each branch, i.e. different branch networks share the same loss function, and the loss function of multiple branches is fused in the residual convolutional network and the back propagation process is performed for network training.

[0086] The positioning accuracy of the method of the application and other visible light positioning methods under different signal-to-noise ratios is compared Figure 4 ​It can be seen that the positioning performance of the application is superior to other positioning methods under different signal-to-noise ratios, and when the signal-to-noise ratio is above 0dB, the positioning accuracy of the application is higher than that of other positioning methods with the increase of the signal-to-noise ratio.

[0087] The above is a specific description of the preferred embodiment of the application, but the application is not limited to the described embodiments, and those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the application, and these equivalent modifications or replacements are all included in the scope defined by the claims of the present application.

Claims

1.A method for visible light indoor positioning based on multi-branch residual convolutional network, characterized in that, The method comprises the following steps: obtaining a visible light RSS data graph and constructing a training set; training a preset positioning model based on the training set to obtain a trained positioning model; the preset positioning model comprises a residual convolutional network and a multi-branch feature fusion network; obtaining a to-be-tested RSS data graph and inputting the RSS data graph into the trained network model to obtain corresponding position information; the step of training the preset positioning model based on the training set to obtain the trained positioning model specifically comprises: training the preset positioning model based on the training set; inputting the normalized RSS data graph into the residual convolutional network to extract feature information graphs at different levels; sending the feature information graphs at different levels into the multi-branch feature fusion network respectively to fuse the feature texture information at the shallow and deep levels to obtain position estimation information; adjusting parameters of the residual convolutional network and the multi-branch feature fusion network in combination with real position labels in the training set to obtain the trained positioning model; the residual convolutional network comprises an initial convolutional layer and a residual learning module, and the step of inputting the normalized RSS data graph into the residual convolutional network to extract feature information graphs at different levels specifically comprises: inputting the normalized RSS data graph into the residual convolutional network; extracting primary texture information of the RSS data graph based on the initial convolutional layer; learning residual information between input features and output features based on the residual learning module connected by a shortcut; extracting high-level features based on the residual learning module and transmitting the high-level features to a next-level residual learning module; in the training process, model parameters associated to different branch networks in the residual convolutional network are fused with gradient information of different branch networks when the model parameters are updated in gradient. An acceptance intensity value formula of each data point of the visible light RSS data graph is as follows: In the above equation, z m,n represents the received intensity value between the nth visible light receiver and the mth LED, Ψ m,n represents the system parameter, r m represents the Lambertian coefficient of the mth LED lamp, x n represents the three-dimensional position of the receiver, p m represents the three-dimensional position of the LED lamp, φ m,n represents the LED emission angle, θ m,n represents the receiver incident light angle, ∈ m,n represents the environmental noise. 2.The visible light indoor positioning method based on the multi-branch residual convolutional network according to claim 1, wherein, the step of obtaining the visible light RSS data graph and constructing the training set specifically comprises: simulating indoor visible light propagation based on a Lambert model under preset conditions; obtaining an RSS data graph under conditions that bulb power, a Lambert coefficient and a fast decay system dynamically change; normalizing the RSS data graph and real position labels and constructing the training set. 3.The method of claim 2, wherein, the preset conditions are: uniformly distributing LED lamps on a roof with a direction vertically downward; randomly distributing positioning targets in a room position and randomly distributing directions; a central visible light receiver has a direction vertically upward to the ground, and the rest of the visible light receivers are uniformly distributed on a circle edge with the central receiver as a core and form a 45-degree angle with the direction of the central receiver. 4.The method of claim 3, wherein, the multi-branch feature fusion network comprises a multi-branch network, a global average pooling layer and a full connection layer, and the step of sending the feature information graphs at different levels into the multi-branch feature fusion network respectively to fuse the feature texture information at the shallow and deep levels to obtain position estimation information specifically comprises: inputting the feature maps mentioned by the plurality of residual convolutional networks into the multi-branch feature fusion network for fusion; extracting and mapping information of the feature maps based on the multi-branch network; performing global average pooling on the feature maps based on the global average pooling layer to obtain feature point data; The feature point data is spliced and stacked based on a full connection layer, and is mapped with the actual position of the target.

Images