A Location Terminal Scene Perception Method Based on MobileNetV3 Network

Abstract

This invention relates to a scene perception method for a positioning terminal based on a MobileNetV3 network, comprising the following steps: Step S1: Acquire fingerprint data in various working scenarios and construct a fingerprint database; Step S2: Perform dimensionality reduction and scene annotation on the fingerprint data in the fingerprint database to obtain processed fingerprint data; Step S3: Construct an improved MobileNetV3 network model and train it based on the processed fingerprint data to obtain a trained model; Step S4: Perform model pruning, quantization, and knowledge distillation on the trained network model to obtain an optimal model; Step S5: Perform scene perception on the collected fingerprint data based on the optimal model to achieve scene classification. This invention achieves low power consumption and high efficiency in scene perception while ensuring accuracy in complex scene perception.

Description

Technical Field

[0001] This invention relates to the field of terminal positioning, and specifically to a positioning terminal scene perception method based on the MobileNetV3 network. Background Technology

[0002] The positioning scenarios for terminal devices are divided into indoor and outdoor scenarios. Indoor scenarios include simple indoor scenarios (line-of-sight communication, weak multipath effects) and complex indoor scenarios (non-line-of-sight communication, strong multipath effects). Outdoor scenarios include open outdoor scenarios, bridge and tunnel scenarios, and rainy / snowy scenarios. Due to the long distance, large fading, and high dynamic range of satellite signal transmission, the BeiDou positioning system is only suitable for outdoor scenarios. Indoors, it needs to rely on external signal sources (including wireless communication technologies such as 5G, Wi-Fi, Bluetooth, and cellular mobile networks) and utilize positioning algorithms to achieve indoor positioning. Due to the complex working environment of terminals, a single positioning algorithm cannot meet the requirements for seamless and accurate indoor and outdoor positioning in all weather conditions and scenarios. Therefore, scene perception of the terminal's environment is required to switch to a specific positioning technology based on the specific scenario. The accuracy of scene perception may directly affect the selection of positioning algorithms, thereby interfering with positioning accuracy. Therefore, achieving scene perception, i.e., classifying the working environment in which the terminal is located, has become an important research problem in positioning tasks. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a scene perception method for positioning terminals based on the MobileNetV3 network, which ensures the accuracy of scene perception in complex scenes while achieving low power consumption and high efficiency in scene perception.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A scene perception method for a positioning terminal based on MobileNetV3 network includes the following steps:

[0006] Step S1: Obtain fingerprint data from various work scenarios and construct a fingerprint database;

[0007] Step S2: Perform dimensionality reduction and scene annotation on the fingerprint data in the fingerprint database to obtain the processed fingerprint data;

[0008] Step S3: Construct an improved MobileNetV3 network model and train it based on the processed fingerprint data to obtain the trained model.

[0009] Step S4: Perform model pruning, quantization, and knowledge distillation on the trained network model to obtain the optimal model;

[0010] Step S5: Based on the optimal model, perform scene perception on the collected fingerprint data to achieve scene classification.

[0011] Furthermore, the fingerprint data acquisition in step S1 is specifically as follows: all information sent by surrounding APs is collected by the terminal, and then the fingerprint data is calculated by means of the filtering method. The fingerprint data consists of the RSS received signal strength from the base station.

[0012] Furthermore, the scene annotations are divided according to actual working scenarios and needs, specifically into two superclasses: indoor scenes and outdoor scenes. Indoor scenes are further divided into two subclasses: simple indoor scenes and complex indoor scenes. Outdoor communication is divided into three subclasses: open outdoor scenes, bridge and tunnel scenes, and rain and snow weather scenes.

[0013] Furthermore, the improved MobileNetV3 network model, based on MobileNetV3, transforms all two-dimensional convolutions in MobileNetV3 into one-dimensional convolutions, discards the Inverted Residual Block with a stride of 2 in MobileNetV3, and only uses the Inverted Residual Block with a stride of 1, and replaces AveragePooling in the MobileNetV3 network with Adaptive Average Pooling.

[0014] Furthermore, step S4 specifically includes:

[0015] Step S41: Add sparsity constraints to each channel of the convolutional kernel in the network, and set all parameters with the smallest sparsity constraints to 0;

[0016] Step S42: Change the data type of the convolution kernels in the network from float to int to reduce the model memory size.

[0017] Furthermore, step S5 also includes converting the optimal model written using the PyTorch framework into ONNX format, deploying it to the positioning terminal, deploying the positioning terminal to the actual working scene, collecting fingerprint data, and performing scene perception on the collected fingerprint data.

[0018] Furthermore, the positioning terminal will employ different positioning algorithms based on the detected scene category to improve positioning accuracy.

[0019] Furthermore, UWB, AOA, and TDOA positioning technologies are used in indoor scenarios, while BeiDou positioning technology is used in outdoor scenarios.

[0020] Compared with the prior art, the present invention has the following advantages:

[0021] This invention ensures the accuracy of perception in complex scenes while achieving low power consumption and high efficiency in scene perception. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention;

[0023] Figure 2 This is an Inverted Residual Block structure in one embodiment of the present invention;

[0024] Figure 3 This is a schematic diagram of adding an SE module to an nverted Residual Block in one embodiment of the present invention;

[0025] Figure 4 This is a diagram of the MobileNetV3 network structure of this invention;

[0026] Figure 5 This invention improves the Inverted Residual Block structure. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] Please refer to Figure 1 This invention provides a method for scene perception of a positioning terminal based on the MobileNetV3 network, comprising the following steps:

[0029] Step S1: Obtain fingerprint data from various work scenarios and construct a fingerprint database;

[0030] Step S2: Perform dimensionality reduction and scene annotation on the fingerprint data in the fingerprint database to obtain the processed fingerprint data;

[0031] Step S3: Construct an improved MobileNetV3 network model and train it based on the processed fingerprint data to obtain the trained model.

[0032] Step S4: Perform model pruning, quantization, and knowledge distillation on the trained network model to obtain the optimal model;

[0033] Step S5: Based on the optimal model, perform scene perception on the collected fingerprint data to achieve scene classification.

[0034] In this embodiment, the fingerprint data acquisition in step S1 is specifically as follows: the terminal collects all information sent by surrounding APs, and then calculates the fingerprint data by means of the filtering method. The fingerprint data consists of the RSS received signal strength from the base station.

[0035] Fingerprint data can be obtained through the following three methods:

[0036] 1. Extensive manual collection of fingerprint data by technicians has the advantage of standardized data collection, fewer outliers and anomalies, and stable data due to their sophisticated equipment and standardized data collection operations. However, it is time-consuming and labor-intensive, and it is difficult to quickly build a fingerprint database.

[0037] 2. Using crowdsourcing to assign fingerprint data collection tasks to non-professionals can quickly collect fingerprint information, but due to the specialization of the process, some anomalies may be collected, leading to contamination of the fingerprint database.

[0038] 3. Utilizing the concept of mobile crowdsourcing and leveraging the mobility of a large number of ordinary users, an automatic wireless signal fingerprint map construction technology was designed. This technology transforms on-site surveys, which previously required specialized equipment and specialized tasks, into a process unconsciously completed by a large number of ordinary users using ordinary devices. This successfully eliminates the time-consuming and labor-intensive manual on-site survey step in fingerprint localization, significantly reducing the deployment cost of fingerprint localization methods. Real-world experiments verified that the automatically constructed fingerprint map performs comparably to manual surveys. A fingerprint database was built by collecting a large amount of fingerprint data in various work scenarios.

[0039] In this embodiment, scene labeling is divided according to actual working scenarios and requirements, specifically into two superclasses: indoor scenes and outdoor scenes. Indoor scenes are further divided into two subclasses: simple indoor scenes (line-of-sight communication, weak multipath effect scenarios) and complex indoor scenes (non-line-of-sight communication, strong multipath effect scenarios). Outdoor communication is divided into three subclasses: open outdoor scenes, bridge and tunnel scenes, and rain and snow weather scenes.

[0040] In this embodiment, the MobileNetV3 network adopts the idea of ​​separable convolution and the Residual Block structure design concept of the ResNet network. The network consists of several Inverted Residual Blocks. Each Inverted Residual Block contains a Projection Layer, a Depthwise convolutional layer, an Expansion Layer, and an H-swish activation function, and retains the shortcut mechanism in the Residual Block. The Inverted Residual Block structure is as follows: Figure 2 As shown.

[0041] Among them, the Projection Layer is a 1*1 convolution with channel dimensionality upscaling; the Depthwise convolutional layer is a 3*3 grouped convolution, which groups each channel; and the Expansion Layer is a 1*1 convolution with channel dimensionality downscaling.

[0042] The inverted Residual Block forms a spindle-shaped structure, ensuring that sufficient features are preserved when using Depthwise convolution for feature extraction.

[0043] The MobileNetV3 network also incorporates SE structures into some Inverted Residual Blocks to construct the internal dependencies between channels of the feature maps and adjust the features accordingly, thus improving network accuracy. Examples of Inverted Residual Blocks with SE structures include... Figure 3 As shown.

[0044] MobileNetV3 utilizes two AutoML techniques to find the optimal neural network architecture for a given problem. First, it uses MnasNet for a coarse-grained structure search, then reinforcement learning selects the optimal configuration from a set of discrete choices. Afterward, MobileNetV3 uses NetAdapt to fine-tune the architecture, demonstrating NetAdapt's complementary capabilities by adjusting underutilized activation channels with small reductions. The MobileNetV3 network architecture diagram is shown below. Figure 4 As shown.

[0045] In this embodiment, step S4 specifically includes:

[0046] Step S41: Add sparsity constraints to each channel of the convolutional kernel in the network, and set all parameters with the smallest sparsity constraints to 0 to achieve model pruning.

[0047] Step S42: Change the data type of the convolution kernel in the network from float to int to reduce the model memory size, thereby achieving quantization and further reducing model data storage.

[0048] In this embodiment, step S5 further includes converting the optimal model written using the PyTorch framework into ONNX format, deploying it to the positioning terminal, deploying the positioning terminal to the actual working scene, collecting fingerprint data, and performing scene perception on the collected fingerprint data.

[0049] Specifically, since mainstream deep learning frameworks now require Python to be written, but terminal devices usually only support C or C++, if you want to deploy neural networks to terminal devices, you can use the ONNX format to convert the Python-written neural network into a C++ framework.

[0050] Open Neural Network Exchange (ONNX) is an open format proposed by Microsoft for simply representing deep learning models. Mainstream deep learning frameworks (including PyTorch, TensorFlow, and Keras) can export or convert to the standard ONNX format, enabling models to be transferred between different frameworks and run on various platforms and terminal devices.

[0051] In this embodiment, the positioning terminal employs different positioning algorithms based on the detected scene category to improve positioning accuracy. UWB, AOA, and TDOA positioning technologies are used for indoor scenes, while BeiDou positioning technology is used for outdoor scenes. Different positioning algorithms can also be used as needed for simple indoor scenes, complex indoor scenes, open outdoor scenes, bridge and tunnel scenes, and rain / snow weather scenes to achieve even higher positioning accuracy.

[0052] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made within the scope of the claims of the present invention should be included in the scope of the present invention.

Claims

1. A positioning terminal scene perception method based on a MobileNetV3 network, characterized in that, The method comprises the following steps: Step S1: acquiring fingerprint data in each working scene and constructing a fingerprint database; Step S2: performing data dimension reduction and scene labeling on the fingerprint data in the fingerprint database to obtain processed fingerprint data; Step S3: constructing an improved MobileNetV3 network model, and training the model based on the processed fingerprint data to obtain a trained model; Step S4: performing model pruning, quantization and knowledge distillation on the trained network model to obtain an optimal model; Step S5: performing scene perception on the collected fingerprint data based on the optimal model to realize scene classification. The improved MobileNetV3 network model is based on MobileNetV3, converts all two-dimensional convolutions in MobileNetV3 into one-dimensional convolutions, discards the Inverted Residual Block with a step of 2 in MobileNetV3, and only selects the Inverted Residual Block with a step of 1, and replaces the Average Pooling in the MobileNetV3 network with the AdaptiveAverage Pooling.

2. The MobileNetV3 network-based positioning terminal scene awareness method according to claim 1, characterized in that, In the step S1, the fingerprint data is acquired as follows: all information sent by surrounding APs is collected through a terminal, and the fingerprint data composed of RSS received signal strengths from base stations is calculated through a filtering method and an average. 3.The MobileNetV3 network-based positioning terminal scene awareness method according to claim 1, wherein, The scene labeling is divided according to actual working scenes and requirements, and specifically divided into two superclasses, including an indoor scene and an outdoor scene, the indoor scene is divided into two subclasses, including a simple indoor scene and a complex indoor scene, and the outdoor scene is divided into three subclasses, including an open outdoor scene, a bridge and tunnel scene and a rainy and snowy weather scene. 4.The MobileNetV3 network-based positioning terminal scene awareness method according to claim 1, wherein, The step S4 specifically comprises: Step S41: adding a sparse constraint to each channel of a convolution kernel in the network, and setting all parameter results with the smallest sparse constraint to 0; Step S42: converting the data type of the convolution kernel in the network from float to int to reduce the model memory size. 5.The MobileNetV3 network-based positioning terminal scene awareness method according to claim 1, wherein, The step S5 further comprises converting the optimal model written by using a Pytorch framework into an ONNX format, deploying the model into a positioning terminal, deploying the positioning terminal into an actual working scene, collecting fingerprint data, and performing scene perception on the collected fingerprint data.

6. The MobileNetV3 network-based positioning terminal scene awareness method according to claim 5, characterized in that, The positioning terminal adopts different positioning algorithms according to the detected scene category to improve the positioning accuracy.

7. The MobileNetV3 network-based positioning terminal scene awareness method according to claim 6, characterized in that, The indoor scene adopts UWB, AOA and TDOA positioning technologies, and the outdoor scene adopts Beidou positioning technology.

Images