Background
In the information age of today, the rise of Mobile Ad hoc Network (MANET) has become a hot topic, and the vehicle Ad hoc Network generated thereby promotes the rapid development of intelligent traffic systems, and meanwhile, as the Mobile Ad hoc Network can provide real-time driving assistance application and road information perception capability for drivers, the efficiency and safety of people going out are improved. The automatic control of the vehicle terminal and the effective perception of the road need the support of the position information, so how to realize the accurate real-time positioning of the vehicle still is a field with important research significance. Currently, known common vehicle positioning methods include a plurality of large-scale and civil global navigation satellite systems represented by global positioning systems, laser ranging radars, vehicle detection and positioning based on image processing, and positioning by extracting Channel State Information (CSI) of wireless signals, and these positioning methods have respective characteristics and also have corresponding disadvantages, and in outdoor positioning, GPS has better positioning accuracy and use cost, but in urban traffic environments, shadows such as greenhouses, buildings, tunnels and the like exist, which easily causes serious accuracy errors and signal loss. Expensive costs due to LiDAR are unacceptable to most individual owners. Although the vehicle detection method based on image processing can be applied to practice, there are technical limitations, such as the detection effect of the vehicle is severely limited by the driving speed and the use environment, including weather factors, so that a new positioning solution is urgently needed to solve the existing problems.
With the rapid development of information technology and computer science, more and more researchers combine signal fingerprints and deep learning algorithms to achieve estimation of target position. The radio fingerprint contains unique signal transmission characteristics that can be used to authenticate a mobile phone or any other radio transceiver terminal. The performance of positioning by fingerprint identification is not easily affected by multipath effect, and generally, the received signal strength and CSI of the signal at the receiving end can represent the signal fingerprint. The deep learning method is different in that a Convolutional Neural Network (CNN) is applied to vehicle position estimation under an outdoor urban traffic environment, and a speed component of a vehicle fused by DOA and TOA contained in CSI of a received signal is used as a training database sample of the CNN.
Disclosure of Invention
The invention aims to provide a deep learning method for realizing the position estimation of a moving target based on multi-path channel state information. The method takes a vehicle mobile terminal as an example, a target node obtains Multipath Channel State Information (MCSI) by using single RSU ranging, and in different driving paths of the vehicle node, the method measures and collects a large number of vehicle receiving signals so as to extract the Direction of Arrival (DOA) and the Time of Arrival (TOA) of a Line of Sight (LOS) path, and fuses the historical speed Information of the vehicle as the fingerprint of the position of the vehicle curtain to train the convolutional neural network. The method has the advantages that a Manhattan moving model is established for running vehicles in the urban road network in the training data collection process, the running paths of the vehicles in the target area are related to the steering probability of road intersections, and the deep learning method can effectively predict the running positions of vehicle terminals through DOA, TOA and speed in the urban traffic environment.
The technical scheme is as follows: in order to solve the technical problem, the invention provides a deep learning method for realizing the estimation of the position of a moving target based on multi-path channel state information, which is suitable for the accurate estimation of the position of a vehicle in an urban environment; the method comprises the following steps of (Roadside Unit, RSU) realizing accurate estimation of vehicle position in urban environment, and (Roadside Unit, RSU) realizing accurate estimation of vehicle position in urban environment by adopting a deep learning method, wherein the deep learning method comprises the following steps: adopting a Convolutional Neural Network (CNN) method, receiving Multipath Channel State Information (MCSI) of signals with target vehicle nodes, deploying roadside units with larger communication radius to cover a corresponding area (block), receiving Multipath signals generated by special Short distance Communications (DSRC) by a vehicle receiving end when a target vehicle runs in the RSU signal coverage area by adopting a uniform linear array antenna, extracting the Arrival angle (Direction of Arrival, DOA) and the Arrival Time (Time of Arrival, TOA) of the received signals on a line-of-sight path, and estimating the position of the next Time by using the CNN according to the current speed Information of the vehicle; in addition, the method adopts a 2D-ESPRIT algorithm to process MCSI so as to obtain effective DOA and TOA positioning parameters; the method comprises the following steps:
(1) the roadside unit with enough communication distance is deployed in a building in a city, and the optimal position of the roadside unit is located at the top end of the building so as to realize the line-of-sight propagation of signals;
(2) when a target vehicle runs in a coverage area, the Manhattan mobile model is applied to vehicle mobility modeling in an urban area, so that the traffic scene of the city can be simulated very truly;
(3) when a mobile node such as an automobile runs on a grid street of the model, the mobile node needs to move along a horizontal street and a vertical street, the running direction of the vehicle at a road intersection is related to the probability, and the probability of turning left or right is generally 0.25 by default, and the probability of going straight is 0.5 by default. In the Manhattan model, the relative motion relation between vehicle nodes has high spatial correlation and time correlation, so that the driving speed of the vehicle has corresponding condition limitation;
(4) generating DOA and TOA data when the vehicle runs by using the vehicle movement model and a 2D-ESPRIT algorithm, and generating a corresponding data set;
(5) fusing the speed component of the vehicle driving direction, and using the three-dimensional data set for training a convolutional neural network to obtain available CNN;
(6) when the convolutional neural network training is finished, when a vehicle runs in a coverage area, the current position information of the vehicle can be predicted by using real-time DOA, TOA and speed data received by vehicle-mounted equipment;
the vehicle position estimation method based on the 2D-ESPRIT algorithm and the convolutional neural network can effectively solve the problem of accurate positioning of the vehicle in the urban environment.
The deployment location of the roadside units is determined by the height of the urban building to ensure that the communication performance of the single RSU deployment is optimized. According to fig. 1, when the deployed position of the RSU is on the top floor of a building, the signal propagation path between the vehicle and the RSU can be considered to have a line-of-sight path, i.e., a large communication radius.
When the vehicle runs in the RSU coverage range, the movement model of the vehicle depends on a Manhattan model, namely, the path of the vehicle running in the target area is related to the turning probability of a road intersection, and the vehicle has to move along a horizontal street and a vertical street, and corresponding condition limits exist for the running speed of the vehicle.
The DOA and TOA parameters on the sight line path of the received signal are obtained by using a 2D-ESPRIT algorithm, a training data set of the CNN is generated according to the speed component of the vehicle and the Manhattan mobile model, and after the CNN training is finished, the vehicle end can receive real-time RSU ranging signals and carry out position estimation on the vehicle through the CNN.
In a vehicle ad hoc network, it is crucial to accurately estimate and track the position of a target vehicle in real time, and the automatic driving or safety application of the vehicle relies on the perception of position information. The deep learning method provides a method for estimating the position of a vehicle by utilizing multi-path channel state information aiming at the defects of a plurality of existing positioning modes, the method utilizes single RSU (remote subscriber Unit) ranging to obtain MCSI (Multi-Carrier-time Signal), and in different driving paths of vehicle nodes, the method measures and collects a large number of vehicle receiving signals so as to extract the arrival angle and the arrival time of a line-of-sight path, integrates the historical speed information of the vehicle and trains a convolutional neural network. The method is characterized in that a Manhattan moving model is established for running vehicles in an urban road network, the running paths of the vehicles in a target area are related to the steering probability of road intersections, and the deep learning method can effectively predict the running positions of vehicle terminals through DOA, TOA and speed in an urban traffic environment;
the deep learning method for realizing vehicle position estimation by using DOA and TOA parameters in multipath channel state information comprises the following steps:
in the vehicle motion model, the motion characteristics and the urban driving environment of the vehicle need to be considered in the vehicle mobility modeling process so as to ensure the authenticity of a simulation result. The Manhattan mobile model is widely applied to modeling of urban areas, and can simulate traffic scenes of downtown areas very truly; when the automobile mobile node runs in a grid street of the model, the automobile mobile node needs to move along a horizontal street and a vertical street; the signal transmission path when the vehicle runs in the coverage area is idealized as LOS (line of sight of wireless transmission signal), and theta and tau represent DOA and TOA of LOS; the driving direction of the vehicle at the road intersection is related to the probability, generally, the probability of turning left or right is defaulted to be 0.25, and the probability of going straight is 0.5; in the manhattan model, the relative motion relation between vehicle nodes has high spatial correlation and time correlation, so that the running speed of the vehicle has corresponding condition limitation.
vi(t+Δt)=vi(t)+Rai(t)Δt (1)
In the above formula vi(t) represents the velocity vector of the ith vehicle (terminal) node at time t, with the value of [ v [ [ v ]min,vmax]The inner parts are uniformly distributed. a isi(t) is an acceleration vector of the node i, and is set to a constant a. R is [ -1, 1 [ ]]Uniformly distributed random numbers. In addition, to ensure a safe distance between the vehicle nodes, the instant speed of the front node i +1 and the node i at the time t should satisfy the following condition.
vi+1(t)=vi(t)+a/2 (2)
The value of the acceleration vector ai +1(t) of the front node i +1 is a, so that the current speed per hour of the vehicle node is related to the speed at the previous moment and depends on the running speed of the front vehicle in the same lane, and the motion state of the vehicle in urban traffic is intuitively reflected.
The MCSI is related to a Channel Frequency Response (CFR) of a received signal, and receives multipath Channel state information generated by Dedicated Short Range Communications (DSRC) at a vehicle receiving end by using a uniform linear array antenna, and a least square estimation of the multipath Channel state information can be expressed as follows:
in the above formula, m and k are the indexes of the sub-antenna array and the sub-carrier, respectively, and the DOA and TOA of the p-th transmission path are represented by θpAnd τpAnd (4) showing. Gamma rayk,pIt is understood that the generalized stationary gaussian process, independent of the different paths and subcarriers, is related to physical phenomena such as reflections of radio propagation, and the frequency of the kth subcarrier can be expressed as fk=fc+fd+kΔf,fc、fdAnd Δ f denote a carrier frequency, a doppler shift, and a subcarrier frequency interval, respectively. N is a radical ofm,kRepresenting additive white gaussian noise at the mth antenna and the kth subcarrier. In the signal model, the multipath channel between the DSRC transmitter and the vehicle receiving end can be completely characterized by CFR, and the CFR is used for extracting positioning parameters such as DOA, TOA and the like, although the CFR has an error in estimation due to the existence of Gaussian white noise, the DOA and the TOA can still be used as the labels of the positions of the vehicle nodes.
For a training data sample of CNN, the method of the invention adopts the method of fusing wireless positioning parameters DOA and TOA with vehicle Speed components as a data training network, firstly, vehicle nodes extract the DOA and TOA of the current position by using OFDM signals of RSU, secondly, the current Speed information is obtained by a Speed sensor of the vehicle nodes, then the collected data is fused to form DOA and TOA-Speed training data, and sampling is carried out at time intervals Tsampling in the driving process of the vehicle to form a training database, and finally, the position estimation of the vehicle is realized by receiving real-time ranging signals.
With the rapid development of information technology and computer science, more and more researchers combine signal fingerprints and deep learning algorithms to achieve estimation of target position. The radio fingerprint contains unique signal transmission characteristics that can be used to authenticate a mobile phone or any other radio transceiver terminal. The performance of positioning by fingerprint identification is not easily affected by multipath effect, and generally, the received signal strength and CSI of the signal at the receiving end can both represent FP of the signal. Positioning algorithms based on RSS and DL are more applied to indoor positioning scenes, and have better positioning performance and practical value. The positioning problem of the multipath signal is processed by using the CSI or the positioning parameters contained in the CSI in an actual scene. Compared with the RSS-only signal tag, the CSI contains more location information about the receiving party, and although it represents the signal characteristics of the physical layer, it can be acquired by higher layers as the corresponding auxiliary network interface card is equipped and the software and hardware are upgraded.
The invention provides a deep learning method for realizing the estimation of a mobile target position based on multi-path channel state information. The difference of the deep learning method is that CNN is applied to vehicle position estimation under outdoor urban traffic environment, for the training data sample of CNN, the method fuses wireless positioning parameters DOA and TOA with vehicle Speed component as data training network, vehicle node extracts DOA and TOA of current position by using multi-path signal of RSU, obtains current Speed information by self Speed sensor, fuses the collected data to form DOA and TOA-Speed training data, and forms time interval T in the driving process of vehiclesamplingSampling is performed to form a training database.
In a Mobile Ad hoc Network (MANET), real-time accurate position estimation and tracking of a moving target are crucial, and intelligent application of a target node depends on sensing of position information. The invention provides a deep learning method for realizing the estimation of a mobile target position by utilizing multi-path Channel State Information aiming at the defects of a plurality of existing mobile positioning modes, which utilizes the ranging of a single infrastructure unit anchor node to enable a target to acquire the multi-path Channel State Information (MCSI), measures and collects a large number of target receiving signals in different motion paths of the target node so as to extract the Arrival angle (Direction of Arrival, DOA) and the Arrival Time (Time of Arrival, TOA) of a Line of Sight (LOS), and fuses the historical speed Information of the target node to train a convolutional neural network. Taking vehicle nodes as an example, a Manhattan movement model is established for running vehicles in an urban road network in the training data collection process, the running paths of the vehicles in a target area are related to the steering probability of road intersections, and the deep learning method can effectively predict the running positions of vehicle terminals through DOA, TOA and speed in an urban traffic environment.
Has the advantages that: the invention utilizes DOA and TOA parameters in CNN and MCSI and speed component of vehicle as label to estimate vehicle position in city environment. Compared with the vehicle detection method based on image recognition, the method has the obvious advantages that the influence of vehicle speed and weather factors can be overcome, the vehicle running in a large range can be predicted, the perception capability of the RSU to vehicle nodes in a coverage area under the multipath environment is improved, and the method has high practical application value.
Detailed Description
As shown in the figure, the deep learning method for realizing accurate estimation of the vehicle position through the RSU in the urban environment adopts the convolutional neural network and the MCSI of the vehicle receiving signal, and is characterized in that: deploying roadside units with larger communication radius to cover corresponding blocks, extracting DOA and TOA of received signals on a line-of-sight path when a target vehicle runs in the RSU signal coverage area, and estimating the position at the next moment by using CNN according to the current speed information of the vehicle; in addition, the method adopts a 2D-ESPRIT algorithm to process MCSI so as to obtain effective DOA and TOA positioning parameters;
the method comprises the following steps: (1) the roadside unit with enough communication distance is deployed in a building in a city, and the optimal position of the roadside unit is located at the top end of the building so as to realize the line-of-sight propagation of signals;
(2) when a target vehicle runs in a coverage area, the Manhattan mobile model is applied to vehicle mobility modeling in an urban area, so that the traffic scene of the city can be simulated very truly;
(3) when a mobile node such as an automobile runs on a grid street of the model, the mobile node needs to move along a horizontal street and a vertical street, the running direction of the vehicle at a road intersection is related to the probability, and the probability of turning left or right is generally 0.25 by default, and the probability of going straight is 0.5 by default. In the Manhattan model, the relative motion relation between vehicle nodes has high spatial correlation and time correlation, so that the driving speed of the vehicle has corresponding condition limitation;
(4) generating DOA and TOA data when the vehicle runs by using the vehicle movement model and a 2D-ESPRIT algorithm, and generating a corresponding data set;
(5) fusing the speed component of the vehicle driving direction, and using the three-dimensional data set for training a convolutional neural network to obtain available CNN;
(6) when the convolutional neural network training is finished, when a vehicle runs in a coverage area, the current position information of the vehicle can be predicted by using real-time DOA, TOA and speed data received by vehicle-mounted equipment;
the vehicle position estimation method based on the 2D-ESPRIT algorithm and the convolutional neural network can effectively solve the problem of accurate positioning of the vehicle in the urban environment.
In the vehicle motion model, the motion characteristics and the urban driving environment of the vehicle need to be considered in the vehicle mobility modeling process so as to ensure the authenticity of a simulation result. The Manhattan mobile model is widely applied to modeling of urban areas, and can simulate traffic scenes of downtown areas very truly. When a mobile node such as an automobile runs in a grid street of the model, the mobile node needs to move along a horizontal street and a vertical street. The signal transmission path when the vehicle is traveling in the coverage area can be idealized as LOS, and θ and τ represent DOA and TOA of LOS. The traveling direction of the vehicle at the road intersection is related to the probability, and generally, the probability of turning left or right is 0.25 by default, and the probability of going straight is 0.5 by default. In the manhattan model, the relative motion relation between vehicle nodes has high spatial correlation and time correlation, so that the running speed of the vehicle has corresponding condition limitation.
vi(t+Δt)=vi(t)+Rai(t)Δt (1)
In the above formula vi(t) represents the velocity vector of the ith vehicle node at time t, with a value of [ v [ [ v ]min,vmax]The inner parts are uniformly distributed. a isi(t) is an acceleration vector of the node i, and may be a constant a. R is [ -1, 1 [ ]]Uniformly distributed random numbers. In addition, to ensure a safe distance between the vehicle nodes, the instant speed of the front node i +1 and the node i at the time t should satisfy the following condition.
vi+1(t)=vi(t)+a/2 (2)
The value of the acceleration vector ai +1(t) of the front node i +1 can be made to be a, so that the current speed per hour of the vehicle node is not only related to the speed at the previous moment, but also depends on the running speed of the front vehicle in the same lane, and the motion state of the vehicle in urban traffic is intuitively reflected.
The MCSI is related to the Channel Frequency Response (CFR) of the received signal, and receives a multipath signal generated by Dedicated Short Range Communications (DSRC) at the vehicle receiving end by using a uniform linear array antenna, and the least square estimation of the multipath signal can be expressed as follows.
In the above formula, m and k are the indexes of the sub-antenna array and the sub-carrier, respectively, and the DOA and TOA of the p-th transmission path are represented by θpAnd τpAnd (4) showing. Gamma rayk,pIt can be understood as a generalized stationary gaussian process, independent of different paths and subcarriers, related to physical phenomena such as reflections of radio propagation, and the frequency of the kth subcarrier can be expressed as fk=fc+fd+kΔf,fc、fdAnd Δ f denote a carrier frequency, a doppler shift, and a subcarrier frequency interval, respectively. N is a radical ofm,kRepresenting additive white gaussian noise at the mth antenna and the kth subcarrier. In the signal model, the multipath channel between the DSRC transmitter and the vehicle receiving end can be completely characterized by CFR, and the CFR is used for extracting positioning parameters such as DOA, TOA and the like, although the CFR has an error in estimation due to the existence of Gaussian white noise, the DOA and the TOA can still be used as the labels of the positions of the vehicle nodes.
For a training data sample of CNN, the method adopts the method of fusing wireless positioning parameters DOA and TOA with a vehicle Speed component as a data training network, firstly, a vehicle node extracts the DOA and TOA of the current position by using an OFDM signal of an RSU, secondly, the current Speed information is obtained through a Speed sensor of the vehicle node, then the collected data is fused to form DOA and TOA-Speed training data, and sampling is carried out at a time interval Tsampling in the driving process of the vehicle to form a training database, and finally, the position estimation of the vehicle is realized by receiving a real-time ranging signal.
In the fully-connected neural network structure, the final result of the output layer is related to the output of each neuron in the hidden layer, and the forward propagation algorithm using the neuron output of the previous layer as the input of the neuron of the next layer is followed, the output of the neuron comprises a linear part and a nonlinear activation part, and the calculation process is as follows, wherein the connection weight of the ith neuron of the k layer and the jth neuron of the k +1 layer is defined as w in the linear partk+1 jiSo that the matrix operation does not need transposition, the bias of the ith neuron of the k layer can be expressed as bk i. The activation function of a neuron is also defined herein as a (#), so that the j-th neuron in the k + 1-th layer activates the output expression yk+1 jCan be expressed as
Currently, commonly used activation functions for the classification problem include ELU functions, tanh functions, rectified Linear units (ReLU), and sigmod functions, and generally only one activation function is required for each neuron. The hidden layer number of the neural network determines the complexity of the function which can be approximated [8], and the more the layer number, the more the number of network linear parameters, namely the weights w and the bias b, is, so that the parameter tuning of the neural network is a difficult problem. In practice, a gradient descent based approach may be used to update the network parameters using the output layer penalty function and the back propagation algorithm, but this requires that the sample data be supervised during the training process. The loss function of the output layer can generally adopt cross entropy loss or mean square loss, and in the multi-classification problem, the output value can be normalized into a probability value by using a softmax function.
However, the more complex fully-connected neural networks can contain more parameters, so that strong computing power is needed to realize forward propagation and backward propagation, an overfitting phenomenon occurs when the number of sample data sets is not enough to train the network, and in addition, it is very difficult to maintain a huge sample data base and the complex fully-connected neural networks no matter at a vehicle node or at an RSU end, and a common solution to the problem is to perform local connection on the neural networks and share the weight of network parameters, compared with other neural networks, the CNN has two obvious characteristics. CNN is often applied to processing three-dimensional data such as pictures in computer vision, and has a hierarchical structure of an input layer, a convolutional layer (Conv), an activation layer (ReLU), a pooling layer (Pool), a full connection layer (Fc), and an output layer, sample data needs to be equalized and normalized before being input to a neural network, the convolutional layer performs feature extraction by using convolutional cores, feature maps corresponding to different convolutional cores represent features of a certain dimension, and the feature extraction becomes more and more specific as the convolutional layer increases. The excitation layer performs nonlinear mapping on the linear calculation result of the convolution layer. The pooling layer enables dimensionality reduction of the feature map, but can retain large-scale important information. The fully-connected layer acts as a classifier, and then the Softmax function normalizes the classification results to probabilities.
The above description is only a preferred embodiment of the present invention, and the scope of the present invention is not limited to the above embodiment, but equivalent modifications or changes made by those skilled in the art according to the present disclosure should be included in the scope of the present invention as set forth in the appended claims.