WO2020220439A1 - 基于深度神经网络的高速公路交通流量状态识别方法 - Google Patents

基于深度神经网络的高速公路交通流量状态识别方法 Download PDF

Info

Publication number
WO2020220439A1
WO2020220439A1 PCT/CN2019/090874 CN2019090874W WO2020220439A1 WO 2020220439 A1 WO2020220439 A1 WO 2020220439A1 CN 2019090874 W CN2019090874 W CN 2019090874W WO 2020220439 A1 WO2020220439 A1 WO 2020220439A1
Authority
WO
WIPO (PCT)
Prior art keywords
dnn
training
layer
traffic flow
model
Prior art date
Application number
PCT/CN2019/090874
Other languages
English (en)
French (fr)
Other versions
WO2020220439A9 (zh
Inventor
郭军
张小钰
刘晨
高志远
王理庚
李文雨
迟航民
Original Assignee
东北大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 东北大学 filed Critical 东北大学
Publication of WO2020220439A1 publication Critical patent/WO2020220439A1/zh
Publication of WO2020220439A9 publication Critical patent/WO2020220439A9/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24147Distances to closest patterns, e.g. nearest neighbour classification
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/048Activation functions
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent
    • GPHYSICS
    • G08SIGNALLING
    • G08GTRAFFIC CONTROL SYSTEMS
    • G08G1/00Traffic control systems for road vehicles
    • G08G1/01Detecting movement of traffic to be counted or controlled
    • G08G1/0104Measuring and analyzing of parameters relative to traffic conditions
    • G08G1/0125Traffic data processing
    • G08G1/0133Traffic data processing for classifying traffic situation
    • GPHYSICS
    • G08SIGNALLING
    • G08GTRAFFIC CONTROL SYSTEMS
    • G08G1/00Traffic control systems for road vehicles
    • G08G1/065Traffic control systems for road vehicles by counting the vehicles in a section of the road or in a parking area, i.e. comparing incoming count with outgoing count
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02Preprocessing
    • G06F2218/04Denoising
    • G06F2218/06Denoising by applying a scale-space analysis, e.g. using wavelet analysis
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08Feature extraction
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12Classification; Matching

Definitions

  • the invention belongs to the technical field of intelligent transportation, and particularly relates to a method for identifying the state of expressway traffic flow based on a deep neural network.
  • Traffic information plays an important role in traffic management. Inadequate detection of highway traffic information is an important cause of current highway traffic jams and frequent accidents.
  • Currently, most methods of obtaining traffic information are through cameras, which are vulnerable to external light occlusion and high computational cost for image analysis. Therefore, how to obtain real-time traffic status information, be less susceptible to external environmental interference, reduce the computational cost of intelligent transportation, and realize intelligent, efficient, and informatized traffic information detection is an urgent problem in the research of intelligent transportation systems.
  • ITS Intelligent Transportation System
  • An outstanding feature of any modern traffic management system is real-time data collection and online processing of data obtained from sensors. Information from such systems can be used to estimate the state of traffic flow. So far, there are two main methods available for traffic measurement: inductive loop detectors buried under the road and visual analysis systems based on camera installations on the road. Induction loop detectors are reliable for more than 50 years, but are still quite expensive and constrained because they require civil construction and maintenance. Due to the integration requirements of infrastructure and the complexity of video processing, cameras are expensive. In addition, visual data processing systems may fail due to obstructed objects, lighting, and weather conditions.
  • the camera collects on-site image data, analyzes the collected image data through video recognition technology and digital technology, and analyzes to obtain traffic information; however, the system needs to process a large amount of image data, and the stability and accuracy of the calculation results still need to be greatly improved Therefore, there is a high demand for the performance of the computer and the hardware for image processing, which also leads to expensive equipment costs.
  • the technical problem to be solved by the present invention is to provide a method for identifying the state of expressway traffic flow based on a deep neural network in view of the above-mentioned shortcomings of the prior art, so as to realize the recognition of the state of expressway traffic flow.
  • the technical solution adopted by the present invention is: a highway traffic flow state recognition method based on a deep neural network, including the following steps:
  • Step 1 Classify and define the traffic flow state, perform noise reduction processing and feature extraction on the audio signal, use DNN for modeling, and obtain a DNN model for identifying the highway traffic flow state;
  • Step 1.1 Classify and define the traffic flow state of the expressway, and divide the traffic flow state of the expressway into five categories: car-free roads, express roads, normal roads, busy roads, and blocked roads; among them, the speed of the car-free lanes 70Km/h and above, the speed of fast lanes is 60-70Km/h, the speed of normal lanes is 40-60Km/h, the speed of busy lanes is 20-40Km/h, and the speed of blocked lanes is 0-20Km/h ;
  • Step 1.2 Use the audio noise reduction algorithm based on wavelet change to remove background interference and enhance the audio signal of highway traffic;
  • Step 1.3 Using the MFCC feature extraction method based on classical modal decomposition weighting, replace the audio signal of highway traffic with EMD decomposition to obtain the feature parameters of the audio signal MFCC of highway traffic;
  • Step 1.4 Restricted Boltzmann Machine (RBM) is superimposed to form a DNN model, forming a bottom-up discriminative training model, and supervised training is used to reduce the errors between layers from bottom to top Transfer;
  • RBM structure includes visible layer and hidden layer; the visible layer contains random nodes, and the hidden layer contains binary random nodes;
  • Step 1.4.1 Use the extracted audio signal characteristic parameters of the highway traffic flow state as the input data of the DNN model, and use the Gaussian model to model the units in the visible layer of the RBM structure.
  • the hidden layer obeys the Bernoulli distribution and the energy function As shown in the following formula:
  • the parameters x, y, and ⁇ respectively represent the number of visible layers, the number of hidden layers, and the parameter set of a single RBM in the RBM structure;
  • the parameter set ⁇ includes parameters a i , b j , w ij , and a i represents the i- th visible layer Offset, x i represents the visible layer vector, y j represents the hidden layer vector, b j represents the offset of the jth hidden layer, and w ij represents the connection weight of the visible layer i and the hidden layer j;
  • Step 1.4.2 In the DNN model structure, there are connections between the layers of the RBM, but there is no connection within each layer. Under the premise that the visible layer nodes are known, the states of the hidden layer nodes are independent of each other. Under the premise that the nodes of the visible layer are known, the activation probability of the j-th node of the hidden layer is calculated as shown in formula 2. Similarly, under the premise that the nodes of the hidden layer are known, the i-th node of the visible layer is The activation probability calculation is shown in formula 3;
  • Step 1.4.3 In the RBM training process, calculate the maximum log-likelihood function on the training set to obtain the RBM model optimal parameter set ⁇ ′ value, as shown in the following formula:
  • N is the number of training sets
  • Step 1.4.4 Use the Contrastive Divergence (CD) algorithm to estimate the RBM structure, and use the gradient descent algorithm to update the RBM parameters;
  • CD Contrastive Divergence
  • Step 1.4.4.1 Use the training set to initialize the node state of the visible layer of the RBM structure, and calculate the state value of each hidden layer node;
  • Step 1.4.4.2 Use the state value of the hidden layer node to derive the state of the visible layer node and reconstruct the visible layer node;
  • Step 1.4.4.3 After the training is completed, a layer of RBM network structure is obtained, which is used as the input of the next layer of RBM structure, and the contrast divergence algorithm is still used for iterative training to obtain the hidden layer node status, and so on to build all RBM networks structure;
  • Step 2 Pre-training the DNN model for identifying the state of expressway traffic flow
  • Step 2.1 Train multi-layer RBM, and then realize the pre-training of DNN;
  • Step 2.1.1 Given the visible layer vector, calculate the activation vector y of the hidden layer node.
  • the hidden layer vector of the RBM is used as training data to train the RBM of another layer, so the next layer is extracted from the output result of the previous layer.
  • Step 2.1.2 After the RBM stops training, it will obtain the Deep Belief Networks (DBN) with the same number of layers as the RBM.
  • the weight coefficient of each layer of RBM corresponds to the initial value of each layer of DBN, so the Parameters are tuned after DBN initialization;
  • Step 2.2 Use the weight of each layer of DBN as the initial weight of the DNN with the sigmoid function as the activation unit. There are labels in the DNN structure. After the pre-training of the DNN is completed, a layer of randomly initialized softmax is introduced as the output layer , Adjust DNN weight parameters through back propagation algorithm;
  • the input data is the output result vector x l-1 of the upper visible layer, and the nodes in the hidden layer are independent of each other.
  • the calculation formula of conditional probability is shown in formula 5;
  • the calculation formula of the conditional probability of the label h corresponding to the output layer is shown in formula 6;
  • b j represents the offset of the jth hidden layer
  • W l represents the connection weight of the l hidden layer and its corresponding visible layer
  • Step 3 Adjust the parameters of the highway traffic flow state recognition model
  • Step 3.1 After training to obtain DBN and initial weights, use back propagation algorithm to optimize DNN network parameters. In the process of parameter tuning, each frame is marked with its category;
  • Step 3.2 Use the cross-primary objective function to adjust the network parameters to minimize the loss between the training target and the hypothesis;
  • o) the probability that the observed variable o is category i
  • the value of the output vector x L is normalized by the softmax function, and meets the condition And It belongs to a multinomial probability distribution, and its calculation method is shown in the following formula:
  • W is the connection weight between the visible layer and the hidden layer
  • M is the characteristic parameter of the observation vector
  • DNN posterior probability is calculated by the formula derived 7
  • Step 3.3 Based on the DNN training algorithm selected by random data, adjust the mini-batch and learning rate used for each training to reduce the total data volume of the training set;
  • the DNN training algorithm based on random data selection designs three different selection functions, and randomly selects different numbers of training data in the training set according to the selection function as the training subset;
  • T 1 (n) t 1 n ⁇ [0,N],t 1 ⁇ (0,1] (10)
  • N is the total number of iterations in the entire training process
  • T i (k) represents the amount of data selected by the selection function i in the kth iteration
  • the parameters t 1 , t 2 , t 3 Indicates the selected variable
  • c indicates the smallest proportion of the random selected variable
  • Step 4 Use Hidden Markov Model HMM to decode the highway traffic flow state recognition model
  • Step 4.1 Train the highway traffic flow state recognition model based on DNN-HMM
  • Step 4.1.1 Train a state-sharing GMM-HMM highway traffic flow state recognition model, the shared state is determined by the decision tree, and the model obtained after training is set as gmm-hmm;
  • Step 4.1.3 Pre-train the deep neural network of dnn-hmm 1 , and set the deep neural network to dnn pre after training;
  • Step 4.1.4 Use gmm-hmm to arrange the training set, calculate the hidden Markov state corresponding to the training set data, and set the obtained data as H; a stable GMM-HMM model is needed to sort the training set data to obtain Labeled training set data;
  • Step 4.1.5 Adjust the parameters of dnn pre through H, and use the back propagation algorithm to obtain a new deep neural network and set it as dnn new ;
  • Step 4.1.6 Re-estimate the transition probability and observation probability parameters in HMM through dnn-hmm 1 and dnn new , and use the maximum likelihood similarity algorithm to obtain a new deep neural network set to dnn sec ;
  • Step 4.1.7 Rearrange the training set data through dnn new and dnn sec and return to step 4.1.5; exit the algorithm until the accuracy of the result of 4.1.6 does not improve; when the training process does not reach convergence, continue to use DNN-HMM Sort the training set data, and perform iterative training of the deep neural network until the algorithm reaches convergence;
  • Step 4.1.8 Given the training set data, estimate the probability value p( st );
  • Step 4.2 Perform HMM decoding based on the DNN-HMM highway traffic flow state recognition model
  • o is the observation vector s represents the unobservable state sequence in the hidden Markov model
  • T s represents the state marked as s
  • the frame value of T represents the total frame value of the training set
  • Step 5 Use the DNN model to estimate the observation probabilities of the audio signals of different highway traffic flow states, and give the identification results of the highway traffic flow states according to the calculated probability.
  • the beneficial effects of adopting the above technical solutions are: the highway traffic flow state recognition method based on deep neural network provided by the present invention, (1) constructing the highway traffic flow state recognition method based on audio signals, and discussing the model in the recognition method
  • unmanned roads unmanned roads
  • express roads normal roads
  • busy roads and blocked roads.
  • the sound signal is mainly controlled by engine idling noise and horn sound.
  • Analyze from model pre-training, model parameter tuning, model recognition, etc. and determine the selected model and related parameters according to the recognition results and actual application requirements.
  • a DNN training algorithm based on random data selection is proposed.
  • the backpropagation algorithm adjusts model parameters, the amount of training data is reduced and the training time is shortened; a DNN-HMM-based approach is proposed.
  • the highway traffic flow state recognition model training algorithm performs probability calculation to obtain the recognition result. Comparing the model recognition performance of different selection functions in the DNN training algorithm based on random data selection, it can determine the learning rate and attenuation factor value during the optimization and adjustment process of the deep neural network structure parameters; construct DNN-HMM models under different highway traffic conditions, Carry out the verification of the highway traffic flow state recognition model based on DNN-HMM.
  • Based on audio processing it can effectively solve the problems of poor image analysis accuracy and large amount of calculation for dynamic image analysis in the current image analysis technology to detect traffic information.
  • FIG. 1 is a flowchart of a method for identifying a state of expressway traffic flow based on a deep neural network according to an embodiment of the present invention
  • a random selection policy data (T 3) change in performance comparison chart wherein, (A) Embodiment 3 of the present invention is a learning rate at different initial value, the random data selection strategy (T 3 ) Performance change, (b) is the learning rate 1.5, under different attenuation factors, the performance change of the random data selection strategy (T 3 );
  • Fig. 5 is a graph of the recognition rate of expressway traffic flow state based on support vector machine provided by an embodiment of the present invention.
  • the method for identifying the state of highway traffic flow based on a deep neural network includes the following steps:
  • Step 1 Classify and define the traffic flow state, perform noise reduction processing and feature extraction on the audio signal, use DNN for modeling, and obtain a DNN model for identifying the highway traffic flow state;
  • Step 1.1 Classify and define the traffic flow state of the expressway, and divide the traffic flow state of the expressway into five categories: car-free roads, express roads, normal roads, busy roads, and blocked roads; among them, the speed of the car-free lanes 70Km/h and above, the speed of fast lanes is 60-70Km/h, the speed of normal lanes is 40-60Km/h, the speed of busy lanes is 20-40Km/h, and the speed of blocked lanes is 0-20Km/h ;
  • Step 1.2 Use the audio noise reduction algorithm based on wavelet change to remove background interference and enhance the audio signal of highway traffic;
  • Step 1.3 Using the MFCC feature extraction method based on classical modal decomposition weighting, replace the audio signal of highway traffic with EMD decomposition to obtain the feature parameters of the audio signal MFCC of highway traffic;
  • Step 1.4 Restricted Boltzmann Machine (RBM) is superimposed to form a DNN model, forming a bottom-up discriminative training model, and supervised training is used to reduce the errors between layers from bottom to top Transfer;
  • RBM structure includes visible layer and hidden layer; the visible layer contains random nodes, and the hidden layer contains binary random nodes;
  • Step 1.4.1 Use the extracted audio signal characteristic parameters of the highway traffic flow state as the input data of the DNN model, and use the Gaussian model to model the units in the visible layer of the RBM structure.
  • the hidden layer obeys the Bernoulli distribution and the energy function As shown in the following formula:
  • the parameters x, y, and ⁇ respectively represent the number of visible layers, the number of hidden layers, and the parameter set of a single RBM in the RBM structure;
  • the parameter set ⁇ includes parameters a i , b j , w ij , and a i represents the i- th visible layer Offset, x i represents the visible layer vector, y j represents the hidden layer vector, b j represents the offset of the jth hidden layer, and w ij represents the connection weight of the visible layer i and the hidden layer j;
  • Step 1.4.2 In the DNN model structure, there are connections between the layers of the RBM, but there is no connection within each layer. Under the premise that the visible layer nodes are known, the states of the hidden layer nodes are independent of each other. Under the premise that the nodes of the visible layer are known, the activation probability of the j-th node of the hidden layer is calculated as shown in formula 2. Similarly, under the premise that the nodes of the hidden layer are known, the i-th node of the visible layer is The activation probability calculation is shown in formula 3;
  • Step 1.4.3 In the RBM training process, calculate the maximum log-likelihood function on the training set to obtain the RBM model optimal parameter set ⁇ ′ value, as shown in the following formula:
  • N is the number of training sets
  • Step 1.4.4 Use the Contrastive Divergence (CD) algorithm to estimate the RBM structure, and use the gradient descent algorithm to update the RBM parameters;
  • CD Contrastive Divergence
  • Step 1.4.4.1 Use the training set to initialize the node state of the visible layer of the RBM structure, and calculate the state value of each hidden layer node;
  • Step 1.4.4.2 Use the state value of the hidden layer node to derive the state of the visible layer node and reconstruct the visible layer node;
  • Step 1.4.4.3 After the training is completed, a layer of RBM network structure is obtained, which is used as the input of the next layer of RBM structure, and the contrast divergence algorithm is still used for iterative training to obtain the hidden layer node status, and so on to build all RBM networks structure;
  • Step 2 Pre-training the DNN model for identifying the state of expressway traffic flow
  • Step 2.1 Train multi-layer RBM, and then realize the pre-training of DNN;
  • Step 2.1.1 Given the visible layer vector, calculate the activation vector y of the hidden layer node.
  • the hidden layer vector of the RBM is used as training data to train the RBM of another layer, so the next layer is extracted from the output result of the previous layer.
  • Step 2.1.2 After the RBM stops training, it will obtain the Deep Belief Networks (DBN) with the same number of layers as the RBM.
  • the weight coefficient of each layer of RBM corresponds to the initial value of each layer of DBN, so the Parameters are tuned after DBN initialization;
  • Step 2.2 Use the weight of each layer of DBN as the initial weight of the DNN with the sigmoid function as the activation unit. There are labels in the DNN structure. After the pre-training of the DNN is completed, a layer of randomly initialized softmax is introduced as the output layer , Adjust DNN weight parameters through back propagation algorithm;
  • the input data is the output result vector x l-1 of the upper visible layer, and the nodes in the hidden layer are independent of each other.
  • the calculation formula of conditional probability is shown in formula 5;
  • the calculation formula of the conditional probability of the label h corresponding to the output layer is shown in formula 6;
  • b j represents the offset of the jth hidden layer
  • W l represents the connection weight of the l hidden layer and its corresponding visible layer
  • Step 3 Adjust the parameters of the highway traffic flow state recognition model
  • Step 3.1 After training to obtain DBN and initial weights, use back propagation algorithm to optimize DNN network parameters. In the process of parameter tuning, each frame is marked with its category;
  • Step 3.2 Use the cross-primary objective function to adjust the network parameters to minimize the loss between the training target and the hypothesis;
  • o) the probability that the observed variable o is category i
  • the value of the output vector x L is normalized by the softmax function, which meets the condition And It belongs to a multinomial probability distribution, and its calculation method is shown in the following formula:
  • W is the connection weight between the visible layer and the hidden layer
  • M is the characteristic parameter of the observation vector
  • DNN posterior probability is calculated by the formula derived 7
  • Step 3.3 Based on the DNN training algorithm selected by random data, adjust the mini-batch and learning rate used for each training to reduce the total data volume of the training set;
  • the DNN training algorithm based on random data selection designs three different selection functions, and randomly selects different numbers of training data in the training set according to the selection function as the training subset;
  • T 1 (n) t 1 n ⁇ [0,N],t 1 ⁇ (0,1] (10)
  • N is the total number of iterations in the entire training process
  • T i (k) represents the amount of data selected by the selection function i in the kth iteration
  • the parameters t 1 , t 2 , t 3 Indicates the selected variable
  • c indicates the smallest proportion of the random selected variable
  • Step 4 Use Hidden Markov Model HMM to decode the highway traffic flow state recognition model
  • Step 4.1 Train the highway traffic flow state recognition model based on DNN-HMM
  • Step 4.1.1 Train a state-sharing GMM-HMM highway traffic flow state recognition model, the shared state is determined by the decision tree, and the model obtained after training is set as gmm-hmm;
  • Step 4.1.3 Pre-train the deep neural network of dnn-hmm 1 , and set the deep neural network to dnn pre after training;
  • Step 4.1.4 Use gmm-hmm to arrange the training set, calculate the hidden Markov state corresponding to the training set data, and set the obtained data as H; a stable GMM-HMM model is needed to sort the training set data to obtain Labeled training set data;
  • Step 4.1.5 Adjust the parameters of dnn pre through H, and use the back propagation algorithm to obtain a new deep neural network and set it as dnn new ;
  • Step 4.1.6 Re-estimate the transition probability and observation probability parameters in HMM through dnn-hmm 1 and dnn new , and use the maximum likelihood similarity algorithm to obtain a new deep neural network set to dnn sec ;
  • Step 4.1.7 Rearrange the training set data through dnn new and dnn sec and return to step 4.1.5; exit the algorithm until the accuracy of the result of 4.1.6 does not improve; when the training process does not reach convergence, continue to use DNN-HMM Sort the training set data, and perform iterative training of the deep neural network until the algorithm reaches convergence;
  • Step 4.1.8 Given the training set data, estimate the probability value p( st );
  • Step 4.2 Perform HMM decoding based on the DNN-HMM highway traffic flow state recognition model
  • Step 4.2.1 Calculate the probability p(o t
  • o is the observation vector
  • s represents the unobservable state sequence in the hidden Markov model
  • T s represents the state marked as The frame value of s
  • T represents the total frame value of the training set
  • Step 5 Use the DNN model to estimate the observation probabilities of the audio signals of different highway traffic flow states, and give the identification results of the highway traffic flow states according to the calculated probability.
  • MATLAB software is used for experimental simulation.
  • the software runs on HPZ820 workstation.
  • the detailed performance parameters of the workstation are shown in Table 1.
  • the programming simulation software is MATLAB2012 version.
  • the audio data of the experiment is a live recording of outdoor roads.
  • the collection environment is normal weather, excluding rain and snow.
  • the collection time period is between 08:00 and 19:00.
  • These audio data cover car-free roads, express roads, and normal roads. , Busy roads, blocked roads, 5 traffic flow states, data tags are mainly used to manually mark the traffic flow state. All audio data is first uniformly converted into wav format audio with a sampling rate of 48KHz monaural through the audio editing software Cool Edit Pro 2.0.
  • All audio data is divided into two sets, one is the training set and the other is the test set; the training set data is used for model parameter training, and the test set data is used for classification and recognition.
  • the training set data has a total of 400 samples.
  • the number of audio data samples for car-free roads, express roads, normal roads, busy roads, and blocked roads are all 80;
  • the test set data has a total of 200 samples, car-free roads, express roads, and normal
  • the number of audio data samples for roads, busy roads, and blocked roads are all 40.
  • the recognition model gives the recognition result of the audio data to be recognized, and refers to the known category information, determines whether the recognition result of the recognition model is correct, and records the result; finally, the accuracy of the recognition result is counted, and the recognition accuracy is measured by the recognition accuracy ,
  • the recognition accuracy is calculated as follows:
  • P classification accuracy
  • C the number of samples with correct classification results
  • S the total number of samples.
  • the DNN training algorithm based on random data selection determines the usage rate of training data through a selection function.
  • the recognition error rates corresponding to the three different selection functions are shown in Figure 2.
  • the T 1 selection function makes the recognition performance drop greatly; the T 2 selection function also reduces the recognition performance, but it is better than the T 1 selection function; when the data When the utilization rate is 56%, the error rate of the T 3 selection function is at least 25.3%, and the model recognition performance is the best.
  • the T 3 selection function is used in the following experiments.
  • the training process of DNN not only goes through the entire training data set for each iteration, but also constantly changes the learning rate. Therefore, appropriate adjustments to the learning rate affect the entire training process. It is important that a proper learning rate adjustment strategy can make the model reach convergence faster.
  • the learning rate ⁇ of different iteration layers in the training process of back propagation is calculated by the following formula.
  • n is the number of iteration layers
  • K is the total number of iterations
  • M is the initial value of the learning rate
  • l is the layer value when the accuracy of the training set increases below the preset threshold
  • is the attenuation factor
  • This experiment uses a wavelet transform-based denoising algorithm for expressway multi-tone signals for noise reduction, and then uses a classical modal decomposition weighted MFCC feature extraction algorithm for feature parameter extraction.
  • the 13-dimensional MFCC parameters, first-order difference parameters and The second-order difference parameters are combined to form a feature vector, which is used as the input parameter of the highway traffic flow state recognition model based on DNN-HMM.
  • a total of 400 training sample data are selected, and there are 80 audio data for each highway traffic flow state, and the DNN-HMM model library is established respectively.
  • Contrast divergence algorithm is used to construct DNN, DNN model is pre-trained through back propagation algorithm, and DNN training algorithm based on random data selection is proposed to optimize the parameters of back propagation algorithm training DNN model, reduce the number of training samples, and shorten Iteration time: HMM decoding calculation is performed on the DNN model, the probability value of the observation sequence is calculated, and the recognition result is obtained.
  • test data are used for the performance test of the highway traffic flow state recognition model based on deep neural network.
  • the result of the recognition accuracy of highway traffic flow based on DNN-HMM is shown in Figure 4. From Figure 4, it can be seen that the recognition rate of busy roads is lower because the audio signals of busy roads include normal roads.
  • the audio signal also includes the audio signal of the congested road; while the recognition accuracy of the car-free lane and the congested lane is higher, mainly because the characteristics of the two traffic flow audio signals are more distinct.
  • support vector machine is a widely used classifier model.
  • the DNN-HMM-based highway traffic flow state identification model of the present invention and the support vector machine model were compared and analyzed and the experimental results were analyzed.
  • the kernel and kernel functions of the support vector machine are shown in Table 2.
  • the support vector machine models with different kernel functions are used for experiments to determine the support vector machine model. Train the support vector machine models of different kernel functions through the audio signal training set data of the highway traffic flow state, and use the test set data to verify the performance of the trained model.
  • Table 2 The kernel and kernel function of support vector machine
  • the performance information of the DNN-HMM-based highway traffic flow state recognition model and the support vector machine model of different kernel functions of the present invention are summarized, and the summary information is shown in Table 3.
  • Table 3 Through the analysis of the contents of Table 3, it can be seen that the recognition accuracy of each model for busy roads is lower, because the audio signal of busy roads includes the audio signal of normal roads and the audio signal of congested roads; compare and analyze different cores
  • the comprehensive recognition rate of the support vector machine model of the function, the support vector machine whose kernel function is a polynomial has the best performance, and its comprehensive recognition rate is 80.93%; while the DNN-HMM-based highway traffic flow state recognition model of the present invention is The comprehensive recognition rate is 81.058%. Comparing the comprehensive recognition rate of the model in this paper and the performance result of the support vector machine whose kernel function is a polynomial, the model of the present invention has a better effect.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Chemical & Material Sciences (AREA)
  • Analytical Chemistry (AREA)
  • Traffic Control Systems (AREA)
  • Image Analysis (AREA)

Abstract

一种基于深度神经网络的高速公路交通流量状态识别方法,涉及智能交通技术领域。该方法对交通流量状态进行分类并定义,对音频信号进行降噪处理和特征提取,使用深度神经网络DNN进行建模,得到对高速公路交通流量状态进行识别的深度神经网络模型,并对深度神经网络模型进行预训练;然后对深度神经网络模型的参数进行调优;利用隐马尔可夫模型HMM对高速公路交通流量状态识别模型进行解码;最后用深度神经网络模型对不同高速公路交通流量状态的音频信号的观测概率进行估计,根据计算出的概率给出高速公路交通流量状态的识别结果。该方法可以有效解决目前图像分析技术检测交通信息中存在的图像分析准确率欠佳、动态图像分析的计算量大等问题。

Description

基于深度神经网络的高速公路交通流量状态识别方法 技术领域
本发明属于智能交通技术领域,特别涉及一种基于深度神经网络的高速公路交通流量状态识别方法。
背景技术
交通信息在交通管理中发挥着重要作用,高速公路交通信息的检测不力是导致目前高速公路交通拥堵、事故频发的重要原因。目前获取交通信息的方式大多通过摄像机,摄像机存在着易受外部光线遮挡影响和图像分析的计算成本较高的不足。因此,如何实时获取交通状态信息、不易受外界环境干扰、降低智能交通的计算成本,实现交通信息检测的智能化、高效化、信息化,是目前智能交通系统研究中亟待解决的问题。
目前,随着信息技术的快速发展,交通信息的收集在交通情报管理中发挥着越来越重要的作用。公路数量不断增加,乘车出行的人数不断增加,运输系统面临的问题也越来越多。智能交通系统(ITS)作为交通系统的重大项目之一,在交通诱导、交通信号控制、公路管理、电子收费等领域发挥着重要作用。为了拥有一套运行稳定、性能良好的智能交通系统,最基本的是拥有完善的车流量检测和实时信息采集,它用于车辆识别和速度检测等。交通流量检测技术作为智能交通的重要组成部分,广泛应用于收费系统,交通统计等相关工作中。
任何现代交通管理系统的一个突出特点是实时数据采集和从传感器获取数据的在线处理,此类系统的信息可用于交通流量状态的估计上。到目前为止,有两种主要方法可用于交通测量:埋在路面下的感应环路检测器和道路上基于摄像机安装的视觉分析系统。感应回路探测器是可靠的,可达50年以上,但仍然是相当昂贵的和受约束的,因为它们需要土建和维护。由于基础设施的集成要求和视频处理的复杂性,摄像机价格昂贵。此外,由于遮挡物体,照明和天气条件,视觉数据处理系统可能会失败。
通过对高速公路的声音研究,可以从高速公路上获取各种类型的声音,例如车辆轮胎噪声、发动机噪音、喇叭声等,这些声音的累积可以用于监测高速公路的交通状况,有助于解决高高速公路交通流量状态的识别问题。在传统的视频图像识别方法中,存在一些不足:(1)摄像机易受外界环境影响。当外界环境发生变化时,特别是摄像头出现异物遮挡、视频盲区、自然天气的变化,背景环境的亮度会影响摄像机的性能,雨雪天气也会干扰摄像机的图像识别准确率,视频识别甚至可能被更恶劣的条件导致失效。(2)图像处理计算负荷大。摄像机进行现场图像数据采集,将采集的图像数据通过视频识别技术和数字技术进行分析,分析获得交通信息;然而,系统需要处理大量的图像数据,计算结果的稳定性和准确性仍需很大提 高,因此对于计算机和进行图像处理的硬件性能有很高的需求,这也导致了昂贵的设备费用。
为了解决图像分析技术检测交通信息中存在的图像分析准确率欠佳、动态图像分析的计算量大的问题,基于高速公路交通的音频信号的研究,提出一种基于深度神经网络的高速公路交通流量状态识别方法。
发明内容
本发明要解决的技术问题是针对上述现有技术的不足,提供一种基于深度神经网络的高速公路交通流量状态识别方法,实现对高速公路交通流量状态进行识别。
为解决上述技术问题,本发明所采取的技术方案是:基于深度神经网络的高速公路交通流量状态识别方法,包括以下步骤:
步骤1:对交通流量状态进行分类并定义,对音频信号进行降噪处理和特征提取,使用DNN进行建模,得到对高速公路交通流量状态进行识别的DNN模型;
步骤1.1:对高速公路的交通流量状态进行分类并定义,将高速公路的交通流量状态分为五类:无车道路、快速道路、正常道路、繁忙道路、堵塞道路;其中,无车车道的速度为70Km/h及以上,快速车道的速度为60-70Km/h,正常车道的速度为40-60Km/h,繁忙车道的速度为20-40Km/h,堵塞车道的速度为0-20Km/h;
步骤1.2利用基于小波变化的音频降噪算法,去除背景干扰,增强高速公路交通的音频信号;
步骤1.3:利用基于经典模态分解加权的MFCC特征提取方法,对高速公路交通的音频信号用EMD分解代替,求得高速公路交通的音频信号MFCC的特征参数;
步骤1.4:由受限玻尔兹曼机(Restricted Boltzmann Machine,RBM)叠加构成DNN模型,形成一种由下到上的区分性训练模型,采用有监督训练将各层间的误差由下到上传递;RBM结构包括可见层和隐藏层;可见层中包含随机节点,隐藏层中包含二值随机节点;
步骤1.4.1:将提取出的高速公路交通流量状态的音频信号特征参数作为DNN模型的输入数据,使用高斯模型对RBM结构可见层中的单元建模,隐藏层服从伯努利分布,能量函数如下公式所示:
Figure PCTCN2019090874-appb-000001
其中,参数x、y、θ分别代表RBM结构中可见层数目、隐藏层数目、单个RBM的参数集合;参数集合θ包括参数a i,b j,w ij,a i代表第 i个可见层的偏移量,x i表示可见层向量,y j表示隐藏层向量,b j代表第j个隐藏层的偏移量,w ij代表可见层i和隐藏层j的连接权重;
步骤1.4.2:在该DNN模型结构中,RBM的各层之间存在连接,而各层内部无连接,在已知可见层节点的前提下,隐藏层各个节点的状态间是相互独立的,在已知可见层各节点的前提下,隐藏层的第j个节点的激活概率计算如公式2所示;同理,在已知隐藏层各节点的前提下,可见层的第i个节点的激活概率计算如公式3所示;
Figure PCTCN2019090874-appb-000002
Figure PCTCN2019090874-appb-000003
步骤1.4.3:在RBM的训练过程中,计算训练集上最大化对数似然函数,得到RBM模型最优参数集θ′值,如下公式所示:
Figure PCTCN2019090874-appb-000004
其中,N为训练集数目;
步骤1.4.4:采用对比散度算法(Contrastive Divergence,CD)对RBM结构进行估计,使用梯度下降算法进行RBM参数更新;
步骤1.4.4.1:使用训练集初始化RBM结构可见层节点状态,计算各隐藏层节点状态值;
步骤1.4.4.2:再用得到隐藏层节点的状态值反向推导出可见层节点状态,重构可见层节点;
步骤1.4.4.3:训练结束后,得到一层RBM网络结构,将其作为下一层RBM结构的输入,依旧使用对比散度算法迭代训练,得到隐藏层节点状态,以此类推搭建全部的RBM网络结构;
步骤2:对高速公路交通流量状态进行识别的DNN模型进行预训练;
步骤2.1:训练多层的RBM,进而实现对DNN的预训练;
步骤2.1.1:给出可见层向量,计算隐藏层节点的激活向量y,RBM的隐藏层向量又被作为训练数据训练另一层的RBM,因此从前一层的输出结果中提取特征得到下一层RBM的权值;
步骤2.1.2:RBM停止训练后,就获得与RBM相同层数的深度信念神经网络(Deep Belief Networks,DBN),每层RBM的权重系数对应为每层DBN的初始值,因此DBN网络中的参数在DBN初始化后进行调优;
步骤2.2:将每层DBN的权值作为以sigmoid函数为激活单元的DNN的初始权值,DNN 结构中是有标签的,在DNN的预训练完成后,引入一层随机初始化的softmax作为输出层,通过反向传播算法进行DNN权值参数调整;
对于DNN模型中某一个隐藏层0<l<L,输入数据为上一层可见层的输出结果向量x l-1,隐藏层中各节点相互独立,条件概率的计算公式如公式5所示;其输出层对应的标签h的条件概率的计算公式如公式6所示;
P(y l|x l-1)=σ(x l-1·W l+b l)   (5)
P(h|x l-1)=soft max(x L-1·W L+b L)   (6)
其中,b j代表第j个隐藏层的偏移量,W l表示l层隐藏层与其对应可见层的连接权重;
步骤3:对高速公路交通流量状态识别模型的参数进行调优;
步骤3.1:训练得到DBN和初始权值后,采用反向传播算法进行DNN网络参数调优,在参数调优的过程中,每一帧被标记出所属类别;
步骤3.2:采用了交叉嫡目标函数进行网络参数调整,实现训练目标和假设类之间损失的最小化;
步骤3.2.1:对于观测向量为O={o 1,o 2,…,o n},神经元的输出类别为q∈{1,…,C},C=N L为类别数目,即高速公路交通流量状态类别;观测变量o是类别i的概率为P dnn(q|o),即输出层的第i个输出
Figure PCTCN2019090874-appb-000005
的值,输出向量x L通过softmax函数进行归一化处理得到,满足条件
Figure PCTCN2019090874-appb-000006
Figure PCTCN2019090874-appb-000007
属于多项概率分布,其计算方法如下公式所示:
Figure PCTCN2019090874-appb-000008
步骤3.2.2:在确定观测向量O={o 1,o 2,…,o n}后,DNN模型通过前向计算逐层计算得到网络结构的输出,具体由参数θ={W l,b l|0<l<L}计算得出;在高速公路交通流量识别模型中,网络结构优化的目标函数计算如下公式所示,在对目标函数进行求导后,用反向传播算法调整DNN网络结构参数;
Figure PCTCN2019090874-appb-000009
其中,W为可见层与隐藏层的连接权重,M为观测向量的特征参数,
Figure PCTCN2019090874-appb-000010
为观测序列O是类别 i的先验概率,
Figure PCTCN2019090874-appb-000011
是由公式7计算得出的DNN的后验概率,v i是通过指示函数v i=I(c=i) 计算得到,具体计算如下公式所示:
Figure PCTCN2019090874-appb-000012
步骤3.3:基于随机数据选择的DNN训练算法,对每次训练使用的mini-batch和学习率进行调整,以减少训练集的总数据量;
基于随机数据选择的DNN训练算法设计三种不同的选择函数,根据选择函数在训练集中随机选择不同数量的训练数据作为训练子集;
所述3种选择函数如公式10、11、12所示:
T 1(n)=t 1 n∈[0,N],t 1∈(0,1]   (10)
Figure PCTCN2019090874-appb-000013
Figure PCTCN2019090874-appb-000014
在公式10、11、12中,N为整个训练过程中的迭代总次数,T i(k)表示选择函数i在第k次迭代时所选择的数据量,参数t 1、t 2、t 3表示选择变量,c表示随机选择变量中的最小比例量;
步骤4:利用隐马尔可夫模型HMM对高速公路交通流量状态识别模型进行解码;
步骤4.1:基于DNN-HMM对高速公路交通流量状态识别模型进行训练;
步骤4.1.1:训练一个状态共享的GMM-HMM的高速公路交通流量状态识别模型,共享状态由决策树确定,训练后获得的模型设为gmm-hmm;
步骤4.1.2:使用gmm-hmm对隐马尔可夫模型进行初始化,设定λ=(A,B,N)的参数值,A为HMM的转移概率,B为HMM的观测概率,N为HMM的状态,获得基于DNN-HMM的的高速公路交通流量状态识别模型设为dnn-hmm 1
步骤4.1.3:对dnn-hmm 1的深度神经网络进行预训练,训练后获得深度神经网络设为dnn pre
步骤4.1.4:使用gmm-hmm对训练集进行排列出来,计算训练集数据对应的隐马可夫状态,获得数据设为H;其中需要用到一个稳定的GMM-HMM模型进行训练集数据排序,获得有标签的训练集数据;
步骤4.1.5:通过H对dnn pre的参数进行调整,采用反向传播算法,获得新的深度神经网络设为dnn new
步骤4.1.6:通过dnn-hmm 1和dnn new重新估计HMM中的转移概率和观测概率参数,采用最大似然相似算法,获得新的深度神经网络设为dnn sec
步骤4.1.7:通过dnn new和dnn sec对训练集数据重新排列,返回步骤4.1.5;直到4.1.6的结果精度不在提高则退出算法;在训练过程没有达到收敛时,继续使用DNN-HMM对训练集数据排序,深度神经网络进行迭代训练直至算法达到收敛;
步骤4.1.8:给出训练集数据,估计概率值p(s t);
步骤4.2:进行基于DNN-HMM高速公路交通流量状态识别模型中的HMM解码;
计算在s t状态下对应观测向量为o t的概率p(o t|s t),通过公式13进行转化计算:
Figure PCTCN2019090874-appb-000015
其中,o为观测向量s代表隐马尔可夫模型中不可观测的状态序列,p(s)=T s/T代表由训练集得出的不同状态的先验概率,T s代表状态标记为s的帧数值,T代表训练集的总帧数值;
步骤5:用DNN模型对不同高速公路交通流量状态的音频信号的观测概率进行估计,根据计算出的概率给出高速公路交通流量状态的识别结果。
采用上述技术方案所产生的有益效果在于:本发明提供的基于深度神经网络的高速公路交通流量状态识别方法,(1)构建基于音频信号的高速公路交通流量状态识别方法,探讨识别方法中的模型建立问题,对高速公路的交通流量状态进行分类:无人道路,快速道路,正常道路,繁忙道路,堵塞道路。对于交通流量小的情况,车辆倾向于中速到高速行驶。另一方面,对于交通流量大的拥挤道路,声音信号主要由发动机怠速噪音和喇叭声所控制。从模型预训练、模型参数调优、模型识别等进行分析,根据识别结果以及实际应用需求确定选用的模型与相关参数。(2)采用对比散度算法进行DNN预训练,提出基于随机数据选择的DNN训练算法,在反向传播算法调整模型参数时,减少了训练数据数量,缩短了训练时间;提出基于DNN-HMM的高速公路交通流量状态识别模型训练算法进行概率计算,获得识别结果。对比基于随机数据选择的DNN训练算法中不同选择函数的模型识别性能,可确定深度神经网络结构参数优化调节过程中的学习率和衰减因子值;构建不同高速公路流量状态下的DNN-HMM模型,进行基于DNN-HMM的高速公路交通流量状态识别模型的验证。(3)基于音频处理,可以有效解决目前图像分析技术检测交通信息中存在的图像分析准确率欠佳、动态图像分析的计算量大等问题。
附图说明
图1为本发明实施例提供的基于深度神经网络的高速公路交通流量状态识别方法的流程图;
图2为本发明实施例提供的三种不同选择函数在不同数据利用率下性能变化对比图;
图3为本发明实施例提供的在不同初始学习率下,随机数据选择策略(T 3)性能变化对比图,其中,(a)为不同的学习率初始值下,随机数据选择策略(T 3)性能变化,(b)为学习率为1.5,不同衰减因子下,随机数据选择策略(T 3)性能变化;
图4为本发明实施例提供的基于DNN-HMM的高速公路交通流量状态识别模型的准确率曲线图;
图5为本发明实施例提供的基于支持向量机的高速公路交通流量状态的识别率曲线图。
具体实施方式
本实施例中,基于深度神经网络的高速公路交通流量状态识别方法,如图1所示,包括以下步骤:
步骤1:对交通流量状态进行分类并定义,对音频信号进行降噪处理和特征提取,使用DNN进行建模,得到对高速公路交通流量状态进行识别的DNN模型;
步骤1.1:对高速公路的交通流量状态进行分类并定义,将高速公路的交通流量状态分为五类:无车道路、快速道路、正常道路、繁忙道路、堵塞道路;其中,无车车道的速度为70Km/h及以上,快速车道的速度为60-70Km/h,正常车道的速度为40-60Km/h,繁忙车道的速度为20-40Km/h,堵塞车道的速度为0-20Km/h;
步骤1.2利用基于小波变化的音频降噪算法,去除背景干扰,增强高速公路交通的音频信号;
步骤1.3:利用基于经典模态分解加权的MFCC特征提取方法,对高速公路交通的音频信号用EMD分解代替,求得高速公路交通的音频信号MFCC的特征参数;
步骤1.4:由受限玻尔兹曼机(Restricted Boltzmann Machine,RBM)叠加构成DNN模型,形成一种由下到上的区分性训练模型,采用有监督训练将各层间的误差由下到上传递;RBM结构包括可见层和隐藏层;可见层中包含随机节点,隐藏层中包含二值随机节点;
步骤1.4.1:将提取出的高速公路交通流量状态的音频信号特征参数作为DNN模型的输入数据,使用高斯模型对RBM结构可见层中的单元建模,隐藏层服从伯努利分布,能量函数如下公式所示:
Figure PCTCN2019090874-appb-000016
其中,参数x、y、θ分别代表RBM结构中可见层数目、隐藏层数目、单个RBM的参数集合;参数集合θ包括参数a i,b j,w ij,a i代表第 i个可见层的偏移量,x i表示可见层向量,y j表示隐藏层向量,b j代表第j个隐藏层的偏移量,w ij代表可见层i和隐藏层j的连接权重;
步骤1.4.2:在该DNN模型结构中,RBM的各层之间存在连接,而各层内部无连接,在已知可见层节点的前提下,隐藏层各个节点的状态间是相互独立的,在已知可见层各节点的前提下,隐藏层的第j个节点的激活概率计算如公式2所示;同理,在已知隐藏层各节点的前提下,可见层的第i个节点的激活概率计算如公式3所示;
Figure PCTCN2019090874-appb-000017
Figure PCTCN2019090874-appb-000018
步骤1.4.3:在RBM的训练过程中,计算训练集上最大化对数似然函数,得到RBM模型最优参数集θ′值,如下公式所示:
Figure PCTCN2019090874-appb-000019
其中,N为训练集数目;
步骤1.4.4:采用对比散度算法(Contrastive Divergence,CD)对RBM结构进行估计,使用梯度下降算法进行RBM参数更新;
步骤1.4.4.1:使用训练集初始化RBM结构可见层节点状态,计算各隐藏层节点状态值;
步骤1.4.4.2:再用得到隐藏层节点的状态值反向推导出可见层节点状态,重构可见层节点;
步骤1.4.4.3:训练结束后,得到一层RBM网络结构,将其作为下一层RBM结构的输入,依旧使用对比散度算法迭代训练,得到隐藏层节点状态,以此类推搭建全部的RBM网络结构;
步骤2:对高速公路交通流量状态进行识别的DNN模型进行预训练;
步骤2.1:训练多层的RBM,进而实现对DNN的预训练;
步骤2.1.1:给出可见层向量,计算隐藏层节点的激活向量y,RBM的隐藏层向量又被作为训练数据训练另一层的RBM,因此从前一层的输出结果中提取特征得到下一层RBM的权值;
步骤2.1.2:RBM停止训练后,就获得与RBM相同层数的深度信念神经网络(Deep Belief Networks,DBN),每层RBM的权重系数对应为每层DBN的初始值,因此DBN网络中的参数在DBN初始化后进行调优;
步骤2.2:将每层DBN的权值作为以sigmoid函数为激活单元的DNN的初始权值,DNN结构中是有标签的,在DNN的预训练完成后,引入一层随机初始化的softmax作为输出层,通过反向传播算法进行DNN权值参数调整;
对于DNN模型中某一个隐藏层0<l<L,输入数据为上一层可见层的输出结果向量x l-1,隐藏层中各节点相互独立,条件概率的计算公式如公式5所示;其输出层对应的标签h的条件概率的计算公式如公式6所示;
P(y l|x l-1)=σ(x l-1·W l+b l)   (5)
P(h|x l-1)=soft max(x L-1·W L+b L)   (6)
其中,b j代表第j个隐藏层的偏移量,W l表示l层隐藏层与其对应可见层的连接权重;
步骤3:对高速公路交通流量状态识别模型的参数进行调优;
步骤3.1:训练得到DBN和初始权值后,采用反向传播算法进行DNN网络参数调优,在参数调优的过程中,每一帧被标记出所属类别;
步骤3.2:采用了交叉嫡目标函数进行网络参数调整,实现训练目标和假设类之间损失的最小化;
步骤3.2.1:对于观测向量为O={o 1,o 2,…,o n},神经元的输出类别为q∈{1,…,C},C=N L为类别数目,即高速公路交通流量状态类别;观测变量o是类别i的概率为P dnn(q|o),即输出层的第i个输出
Figure PCTCN2019090874-appb-000020
的值,输出向量x L通过softmax函数进行归一化处理得到,满足条件
Figure PCTCN2019090874-appb-000021
Figure PCTCN2019090874-appb-000022
属于多项概率分布,其计算方法如下公式所示:
Figure PCTCN2019090874-appb-000023
步骤3.2.2:在确定观测向量O={o 1,o 2,…,o n}后,DNN模型通过前向计算逐层计算得到网络结构的输出,具体由参数θ={W l,b l|0<l<L}计算得出;在高速公路交通流量识别模型中,网络结构优化的目标函数计算如下公式所示,在对目标函数进行求导后,用反向传播算法调整DNN网络结构参数;
Figure PCTCN2019090874-appb-000024
其中,W为可见层与隐藏层的连接权重,M为观测向量的特征参数,
Figure PCTCN2019090874-appb-000025
为观测序列O是类别 i的先验概率,
Figure PCTCN2019090874-appb-000026
是由公式7计算得出的DNN的后验概率,v i是通过指示函数v i=I(c=i)计算得到,具体计算如下公式所示:
Figure PCTCN2019090874-appb-000027
步骤3.3:基于随机数据选择的DNN训练算法,对每次训练使用的mini-batch和学习率进行调整,以减少训练集的总数据量;
基于随机数据选择的DNN训练算法设计三种不同的选择函数,根据选择函数在训练集中随机选择不同数量的训练数据作为训练子集;
所述3种选择函数如公式10、11、12所示:
T 1(n)=t 1 n∈[0,N],t 1∈(0,1]   (10)
Figure PCTCN2019090874-appb-000028
Figure PCTCN2019090874-appb-000029
在公式10、11、12中,N为整个训练过程中的迭代总次数,T i(k)表示选择函数i在第k次迭代时所选择的数据量,参数t 1、t 2、t 3表示选择变量,c表示随机选择变量中的最小比例量;
步骤4:利用隐马尔可夫模型HMM对高速公路交通流量状态识别模型进行解码;
步骤4.1:基于DNN-HMM对高速公路交通流量状态识别模型进行训练;
步骤4.1.1:训练一个状态共享的GMM-HMM的高速公路交通流量状态识别模型,共享状态由决策树确定,训练后获得的模型设为gmm-hmm;
步骤4.1.2:使用gmm-hmm对隐马尔可夫模型进行初始化,设定λ=(A,B,N)的参数值,A为HMM的转移概率,B为HMM的观测概率,N为HMM的状态,获得基于DNN-HMM的的高速公路交通流量状态识别模型设为dnn-hmm 1
步骤4.1.3:对dnn-hmm 1的深度神经网络进行预训练,训练后获得深度神经网络设为dnn pre
步骤4.1.4:使用gmm-hmm对训练集进行排列出来,计算训练集数据对应的隐马可夫状 态,获得数据设为H;其中需要用到一个稳定的GMM-HMM模型进行训练集数据排序,获得有标签的训练集数据;
步骤4.1.5:通过H对dnn pre的参数进行调整,采用反向传播算法,获得新的深度神经网络设为dnn new
步骤4.1.6:通过dnn-hmm 1和dnn new重新估计HMM中的转移概率和观测概率参数,采用最大似然相似算法,获得新的深度神经网络设为dnn sec
步骤4.1.7:通过dnn new和dnn sec对训练集数据重新排列,返回步骤4.1.5;直到4.1.6的结果精度不在提高则退出算法;在训练过程没有达到收敛时,继续使用DNN-HMM对训练集数据排序,深度神经网络进行迭代训练直至算法达到收敛;
步骤4.1.8:给出训练集数据,估计概率值p(s t);
步骤4.2:进行基于DNN-HMM高速公路交通流量状态识别模型中的HMM解码;
步骤4.2.1:计算在s t状态下对应观测向量为o t的概率p(o t|s t),通过公式13进行转化计算:
Figure PCTCN2019090874-appb-000030
其中,o为观测向量,s代表隐马尔可夫模型中不可观测的状态序列,p(s)=T s/T代表由训练集得出的不同状态的先验概率,T s代表状态标记为s的帧数值,T代表训练集的总帧数值;
步骤5:用DNN模型对不同高速公路交通流量状态的音频信号的观测概率进行估计,根据计算出的概率给出高速公路交通流量状态的识别结果。
本实例使用MATLAB软件进行实验仿真,软件运行在HPZ820工作站上,工作站的性能详细参数如表1所示,编程仿真软件为MATLAB2012版本。实验的音频数据是室外道路实况录制,采集环境为正常天气,不包含雨雪天气,采集时间段为08:00到19:00之间,这些音频数据涵盖了无车道路、快速道路、正常道路、繁忙道路、堵塞道路5种交通流量状态,数据标签主要通过人工进行标注交通流量状态。所有的音频数据都先通过音频编辑软件Cool Edit Pro 2.0统一转换为采样率为48KHz单声道的wav格式的音频。所有的音频数据分为两个集合,一个为训练集,另一个为测试集;训练集数据进行模型参数的训练,测试集数据进行分类识别。训练集数据共有400个样本,无车道路、快速道路、正常道路、繁忙道路、堵塞道路的音频数据样本个数均为80个;测试集数据共有200个样本,无车道路、快速道路、正常道路、繁忙道路、堵塞道路的音频数据样本个数均为40个。
表1 HPZ820服务器性能参数表
性能指标 性能参数
CPU类型 四核至强E5-2603
CPU主频 1.8GHz
CPU线程数 24线程
内存类型 DDR3-1333 ECC
内存容量 8GB
硬盘类型 15000转SAS硬盘
硬盘容量 300GB
带宽 1000Mbps
本实施例中,输入一段类型已知的高速公路交通流量状态的音频数据,进行降噪处理,然后提取其特征向量,构造高速公路交通流量状态的音频信号特征向量集,并将其输入到训练好的基于深度神经网络的高速公路交通流量状态识别模型中。识别模型给出待识别的音频数据的识别结果,并且参考已知的类别信息,确定识别模型的识别结果是否正确,并记录结果;最后,统计识别结果的准确度,识别准确度由识别精度测量,识别精度计算如下如公式所示:
Figure PCTCN2019090874-appb-000031
其中,P代表分类精度,C代表分类结果正确的样本数,S代表总的样本数。
在本发明中,基于随机数据选择的DNN训练算法通过选择函数确定训练数据的使用率,在不同的数据使用率情况下,三种不同选择函数对应的识别错误率如图2所示,从图中分析得出,当数据使用率低于77%时,T 1选择函数使识别性能有很大幅度的下降;T 2选择函数也使识别性能下降,但要优于T 1选择函数;当数据使用率为56%时,T 3选择函数的错误率最低为25.3%,模型识别性能达到最佳,T 3选择函数被用于后面的实验中。
在基于深度神经网络的高速公路交通流量状态识别模型的训练过程中,DNN的训练过程不仅每次迭代要经过整个训练数据集,还要不断改变学习率,因此,适当调整学习率对整个训练过程是重要的,合适的学习率调整策略能使模型更快达到收敛。反向传播的训练过程中不同迭代层数的学习率γ通过下公式计算得到。
Figure PCTCN2019090874-appb-000032
在此公式中,n为迭代层数,K为总的迭代次数,M为学习率的初始值,l为训练集正确率增加低于预先设定的阈值时的层数值,τ为衰减因子。
下面进行学习率对随机数据选择的影响。选择函数为T 3时,不同初始学习率和衰减因子下随机数据选择的性能对比如图3所示。分析图3可知更大的初始学习率以及衰减因子,性能提高相对较小。因为在每层迭代中用到了不同的训练数据,用到的训练数据少于整体的训练数据。为保证整体性能,本实验选择M=1.5和τ=0.7用于后面实验。
本实验使用基于小波变换的高速公路多音频信号降噪算法进行降噪处理,然后使用基于经典模态分解加权的MFCC特征提取算法进行特征参数提取,将13维的MFCC参数、一阶差分参数和二阶差分参数结合形成特征向量,作为基于DNN-HMM的高速公路交通流量状态识别模型的输入参数,神经网络的结构中,输入层数量为1,输入层节点数量为429(39*11=429)个,隐藏数设置为3,每个隐藏层节点数量为1024各个节点,输出层数量为1。
针对不同高速公路交通流量状态的音频信号特征,共选取400个训练样本数据,每种高速公路交通流量状态的音频数据有80个,分别来进行DNN-HMM模型库建立。采用对比散度算法进行DNN的构建,通过反向传播算法进行DNN模型的预训练,并提出基于随机数据选择的DNN训练算法,优化反向传播算法训练DNN模型的参数,减少训练样本数量,缩短迭代时间;对DNN模型进行HMM解码计算,求出观测序列概率值,获得识别结果。
采用200个测试数据用于基于深度神经网络的高速公路交通流量状态识别模型的性能测试,每种交通流量状态音频数据有40个。本实施例中,基于DNN-HMM的高速公路交通流量状态识别准确率结果如图4所示,从图4可知,繁忙道路的识别率偏低一些,因为繁忙道路的音频信号即包含了正常道路的音频信号,也包含了拥堵道路的音频信号;而无车车道和拥堵车道识别准确率偏高一些,主要由于这两种交通流量音频信号的特点更加鲜明。
实验结果表明,本发明提出的方法在识别准确性上表现很好,因此,基于深度神经网络的高速公路交通流量状态识别方法是可行的、有效的。
在目前的交通流量状态的识别研究中,支持向量机是应用较广的一种分类器模型。为研究发明提出的基于深度神经网络的高速公路大型车辆识别方法的优越性,将本发明的基于DNN-HMM的高速公路交通流量状态识别模型与支持向量机模型进行对比实验并分析实验结果。其中,支持向量机的核和核函数如表2所示,分别使用不同核函数的支持向量机模型进行实验,确定支持向量机模型。通过高速公路交通流量状态音频信号训练集数据分别对不同核函数的支持向量机模型进行训练,获得训练好的模型使用测试集数据进行性能验证。
表2支持向量机的核和核函数
Figure PCTCN2019090874-appb-000033
针对5种高速公路交通流量状态的音频信号进行分帧加窗和MFCC特征参数的提取,将提取出的5种高速公路交通流量状态的音频信号的特征向量作为支持向量机的输入分别进行训练,得到高速公路交通流量状态的支持向量机模型。接下来进行支持向量机模型性能验证,分别进行了不同核函数的支持向量机模型的性能实验,采用200个测试数据用于支持向量机模型性能测试,测试结果如图5所示。
本发明的基于DNN-HMM的高速公路交通流量状态识别模型和不同核函数的支持向量机的模型的性能信息进行汇总,汇总信息如表3所示。通过对表3内容分析可知,各个模型对于繁忙道路的识别准确率均偏低一些,因为繁忙道路的音频信号即包含了正常道路的音频信号,也包含了拥堵道路的音频信号;对比分析不同核函数的支持向量机模型的综合识别率,核函数为多项式的支持向量机的性能效果最佳,其综合识别率为80.93%;而本发明的基于DNN-HMM的高速公路交通流量状态识别模型的综合识别率为81.058%,对比本文模型的综合识别率与核函数为多项式的支持向量机的性能结果,本发明的模型效果更佳。
表3不同方法的高速公路交通流量状态的识别率
模型 无车 快速 正常 繁忙 拥堵 综合识别率
线性-SVM 81.76 81.03 79.4 78.5 82.9 80.72
二次方-SVM 78.63 82.6 81.3 79.5 80.1 80.43
多项式-SVM 81.54 80.23 81.96 79.93 81.01 80.93
RBF-SVM 80.7 81.6 78.4 79.5 83 80.64
MPL-SVM 81.86 79.5 78.34 78.98 82.9 80.32
DNN-HMM 82.1 81.37 79.74 78.84 83.24 81.058
最后应说明的是:以上实施例仅用以说明本发明的技术方案,而非对其限制;尽管参照前述实施例对本发明进行了详细的说明,本领域的普通技术人员应当理解:其依然可以对前 述实施例所记载的技术方案进行修改,或者对其中部分或者全部技术特征进行等同替换;而这些修改或者替换,并不使相应技术方案的本质脱离本发明权利要求所限定的范围。

Claims (8)

  1. 一种基于深度神经网络的高速公路交通流量状态识别方法,其特征在于:包括以下步骤:
    步骤1:对交通流量状态进行分类并定义,对音频信号进行降噪处理和特征提取,使用DNN进行建模,得到对高速公路交通流量状态进行识别的DNN模型;
    步骤1.1:对高速公路的交通流量状态进行分类并定义,将高速公路的交通流量状态分为五类:无车道路、快速道路、正常道路、繁忙道路、堵塞道路;其中,无车车道的速度为70Km/h及以上,快速车道的速度为60-70Km/h,正常车道的速度为40-60Km/h,繁忙车道的速度为20-40Km/h,堵塞车道的速度为0-20Km/h;
    步骤1.2利用基于小波变化的音频降噪算法,去除背景干扰,增强高速公路交通的音频信号;
    步骤1.3:利用基于经典模态分解加权的MFCC特征提取方法,对高速公路交通的音频信号用EMD分解代替,求得高速公路交通的音频信号MFCC的特征参数;
    步骤1.4:由受限玻尔兹曼机RBM叠加构成DNN模型,形成一种由下到上的区分性训练模型,采用有监督训练将各层间的误差由下到上传递;RBM结构包括可见层和隐藏层;可见层中包含随机节点,隐藏层中包含二值随机节点;
    步骤2:对高速公路交通流量状态进行识别的DNN模型进行预训练;
    步骤2.1:训练多层的RBM,进而实现对DNN的预训练;
    步骤2.2:将每层DBN的权值作为以sigmoid函数为激活单元的DNN的初始权值,DNN结构中是有标签的,在DNN的预训练完成后,引入一层随机初始化的softmax作为输出层,通过反向传播算法进行DNN权值参数调整;
    步骤3:对高速公路交通流量状态识别模型的参数进行调优;
    步骤3.1:训练得到DBN和初始权值后,采用反向传播算法进行DNN网络参数调优,在参数调优的过程中,每一帧被标记出所属类别;
    步骤3.2:采用了交叉嫡目标函数进行网络参数调整,实现训练目标和假设类之间损失的最小化;
    步骤3.3:基于随机数据选择的DNN训练算法,对每次训练使用的mini-batch和学习率进行调整,以减少训练集的总数据量;
    步骤4:利用隐马尔可夫模型HMM对高速公路交通流量状态识别模型进行解码;
    步骤4.1:基于DNN-HMM对高速公路交通流量状态识别模型进行训练;
    步骤4.2:进行基于DNN-HMM高速公路交通流量状态识别模型中的HMM解码;
    步骤5:用DNN模型对不同高速公路交通流量状态的音频信号的观测概率进行估计,根 据计算出的概率给出高速公路交通流量状态的识别结果。
  2. 根据权利要求1所述的基于深度神经网络的高速公路交通流量状态识别方法,其特征在于:所述步骤1.4的具体方法为:
    步骤1.4.1:将提取出的高速公路交通流量状态的音频信号特征参数作为DNN模型的输入数据,使用高斯模型对RBM结构可见层中的单元建模,隐藏层服从伯努利分布,能量函数如下公式所示:
    Figure PCTCN2019090874-appb-100001
    其中,参数x、y、θ分别代表RBM结构中可见层数目、隐藏层数目、单个RBM的参数集合;参数集合θ包括参数a i,b j,w ij,a i代表第i个可见层的偏移量,x i表示可见层向量,y j表示隐藏层向量,b j代表第j个隐藏层的偏移量,w ij代表可见层i和隐藏层j的连接权重;
    步骤1.4.2:在该DNN模型结构中,RBM的各层之间存在连接,而各层内部无连接,在已知可见层节点的前提下,隐藏层各个节点的状态间是相互独立的,在已知可见层各节点的前提下,隐藏层的第j个节点的激活概率计算如公式2所示;同理,在已知隐藏层各节点的前提下,可见层的第i个节点的激活概率计算如公式3所示;
    Figure PCTCN2019090874-appb-100002
    Figure PCTCN2019090874-appb-100003
    步骤1.4.3:在RBM的训练过程中,计算训练集上最大化对数似然函数,得到RBM模型最优参数集θ′值,如下公式所示:
    Figure PCTCN2019090874-appb-100004
    其中,N为训练集数目;
    步骤1.4.4:采用对比散度算法对RBM结构进行估计,使用梯度下降算法进行RBM参数更新;
    步骤1.4.4.1:使用训练集初始化RBM结构可见层节点状态,计算各隐藏层节点状态值;
    步骤1.4.4.2:再用得到隐藏层节点的状态值反向推导出可见层节点状态,重构可见层节点;
    步骤1.4.4.3:训练结束后,得到一层RBM网络结构,将其作为下一层RBM结构的输入,依旧使用对比散度算法迭代训练,得到隐藏层节点状态,以此类推搭建全部的RBM网络结构。
  3. 根据权利要求2所述的基于深度神经网络的高速公路交通流量状态识别方法,其特征在于:所述步骤2.1的具体方法为:
    步骤2.1.1:给出可见层向量,计算隐藏层节点的激活向量y,RBM的隐藏层向量又被作为训练数据训练另一层的RBM,因此从前一层的输出结果中提取特征得到下一层RBM的权值;
    步骤2.1.2:RBM停止训练后,就获得与RBM相同层数的深度信念神经网络DBN,每层RBM的权重系数对应为每层DBN的初始值,因此DBN网络中的参数在DBN初始化后进行调优。
  4. 根据权利要求3所述的基于深度神经网络的高速公路交通流量状态识别方法,其特征在于:所述步骤2.2的具体方法为:
    对于DNN模型中某一个隐藏层0<l<L,输入数据为上一层可见层的输出结果向量x l-1,隐藏层中各节点相互独立,条件概率的计算公式如公式5所示;其输出层对应的标签h的条件概率的计算公式如公式6所示;
    P(y l|x l-1)=σ(x l-1·W l+b l)        (5)
    P(h|x l-1)=soft max(x L-1·W L+b L)        (6)
    其中,b j代表第j个隐藏层的偏移量,W l表示l层隐藏层与其对应可见层的连接权重。
  5. 根据权利要求4所述的基于深度神经网络的高速公路交通流量状态识别方法,其特征在于:所述步骤3.2的具体方法为:
    步骤3.2.1:对于观测向量为O={o 1,o 2,...,o n},神经元的输出类别为q∈{1,...,C},C=N L为类别数目,即高速公路交通流量状态类别;观测变量o是类别i的概率为P dnn(q|o),即输出层的第i个输出
    Figure PCTCN2019090874-appb-100005
    的值,输出向量x L通过softmax函数进行归一化处理得到,满足条件
    Figure PCTCN2019090874-appb-100006
    Figure PCTCN2019090874-appb-100007
    属于多项概率分布,其计算方法如下公式所示:
    Figure PCTCN2019090874-appb-100008
    步骤3.2.2:在确定观测向量O={o 1,o 2,...,o n}后,DNN模型通过前向计算逐层计算得到 网络结构的输出,具体由参数θ={W l,b l|0<l<L}计算得出;在高速公路交通流量识别模型中,网络结构优化的目标函数计算如下公式所示,在对目标函数进行求导后,用反向传播算法调整DNN网络结构参数;
    Figure PCTCN2019090874-appb-100009
    其中,W为可见层与隐藏层的连接权重,M为观测向量的特征参数,
    Figure PCTCN2019090874-appb-100010
    为观测序列O是类别i的先验概率,
    Figure PCTCN2019090874-appb-100011
    是由公式7计算得出的DNN的后验概率,v i是通过指示函数v i=I(c=i)计算得到,具体计算如下公式所示:
    Figure PCTCN2019090874-appb-100012
  6. 根据权利要求5所述的基于深度神经网络的高速公路交通流量状态识别方法,其特征在于:步骤3.3所述基于随机数据选择的DNN训练算法设计三种不同的选择函数,根据选择函数在训练集中随机选择不同数量的训练数据作为训练子集;
    所述3种选择函数如公式10、11、12所示:
    T 1(n)=t 1 n∈[0,N],t 1∈(0,1]       (10)
    Figure PCTCN2019090874-appb-100013
    Figure PCTCN2019090874-appb-100014
    在公式10、11、12中,N为整个训练过程中的迭代总次数,T i(k)表示选择函数i在第k次迭代时所选择的数据量,参数t 1、t 2、t 3表示选择变量,c表示随机选择变量中的最小比例量。
  7. 根据权利要求6所述的基于深度神经网络的高速公路交通流量状态识别方法,其特征在于:所述步骤4.1的具体方法为:
    步骤4.1.1:训练一个状态共享的GMM-HMM的高速公路交通流量状态识别模型,共享状态由决策树确定,训练后获得的模型设为gmm-hmm;
    步骤4.1.2:使用gmm-hmm对隐马尔可夫模型进行初始化,设定λ=(A,B,N)的参数值,A为HMM的转移概率,B为HMM的观测概率,N为HMM的状态,获得基于DNN-HMM的的高速公路交通流量状态识别模型设为dnn-hmm 1
    步骤4.1.3:对dnn-hmm 1的深度神经网络进行预训练,训练后获得深度神经网络设为dnn pre
    步骤4.1.4:使用gmm-hmm对训练集进行排列出来,计算训练集数据对应的隐马可夫状态,获得数据设为H;其中需要用到一个稳定的GMM-HMM模型进行训练集数据排序,获得有标签的训练集数据;
    步骤4.1.5:通过H对dnn pre的参数进行调整,采用反向传播算法,获得新的深度神经网络设为dnn new
    步骤4.1.6:通过dnn-hmm 1和dnn new重新估计HMM中的转移概率和观测概率参数,采用最大似然相似算法,获得新的深度神经网络设为dnn sec
    步骤4.1.7:通过dnn new和dnn sec对训练集数据重新排列,返回步骤4.1.5;直到4.1.6的结果精度不在提高则退出算法;在训练过程没有达到收敛时,继续使用DNN-HMM对训练集数据排序,深度神经网络进行迭代训练直至算法达到收敛;
    步骤4.1.8:给出训练集数据,估计概率值p(s t)。
  8. 根据权利要求7所述的基于深度神经网络的高速公路交通流量状态识别方法,其特征在于:所述4.2的具体方法为:计算在s t状态下对应观测向量为o t的概率p(o t|s t),通过公式13进行转化计算:
    Figure PCTCN2019090874-appb-100015
    其中,o为观测向量s代表隐马尔可夫模型中不可观测的状态序列,p(s)=T s/T代表由训练集得出的不同状态的先验概率,T s代表状态标记为s的帧数值,T代表训练集的总帧数值。
PCT/CN2019/090874 2019-04-29 2019-06-12 基于深度神经网络的高速公路交通流量状态识别方法 WO2020220439A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910353551.9 2019-04-29
CN201910353551.9A CN110097755B (zh) 2019-04-29 2019-04-29 基于深度神经网络的高速公路交通流量状态识别方法

Publications (2)

Publication Number Publication Date
WO2020220439A1 true WO2020220439A1 (zh) 2020-11-05
WO2020220439A9 WO2020220439A9 (zh) 2021-02-11

Family

ID=67446214

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/090874 WO2020220439A1 (zh) 2019-04-29 2019-06-12 基于深度神经网络的高速公路交通流量状态识别方法

Country Status (2)

Country Link
CN (1) CN110097755B (zh)
WO (1) WO2020220439A1 (zh)

Cited By (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112308214A (zh) * 2020-11-09 2021-02-02 海南核电有限公司 一种用于冷源致灾物灾害预测的深度学习改进算法
CN112507624A (zh) * 2020-12-15 2021-03-16 交通运输部公路科学研究所 一种城际公路出行方式识别模型构建、识别方法及装置
CN112836802A (zh) * 2021-02-03 2021-05-25 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) 一种半监督学习方法、岩性预测方法及存储介质
CN112926273A (zh) * 2021-04-13 2021-06-08 中国人民解放军火箭军工程大学 一种多元退化设备剩余寿命预测方法
CN113095381A (zh) * 2021-03-29 2021-07-09 西安交通大学 一种基于改进型dbn的水声目标辨识方法及系统
CN113096413A (zh) * 2021-03-23 2021-07-09 北京交通大学 基于高速公路收费数据的省际出行流量估算方法
CN113240029A (zh) * 2021-05-25 2021-08-10 哈尔滨工程大学 一种gmm-hmm混合模型概率似然得分的svm二次分类方法
CN113537580A (zh) * 2021-06-28 2021-10-22 中科领航智能科技(苏州)有限公司 一种基于自适应图学习的公共交通客流预测方法及系统
CN113612786A (zh) * 2021-08-09 2021-11-05 上海交通大学宁波人工智能研究院 一种车辆总线的入侵检测系统及方法
CN113807404A (zh) * 2021-08-24 2021-12-17 中通服咨询设计研究院有限公司 面向多源信息融合的智慧交通网络智能化监控与应急系统
CN114743379A (zh) * 2022-06-13 2022-07-12 广东邦盛北斗科技股份公司 基于北斗的城市大面积路网交通感知方法、系统及云平台
CN115754010A (zh) * 2022-11-10 2023-03-07 四川科力特硬质合金股份有限公司 机械密封环密封端面质量控制方法
CN116016065A (zh) * 2023-03-24 2023-04-25 南京大学 基于神经网络的抗串扰高速互联信号接收方法及系统
CN116050672A (zh) * 2023-03-31 2023-05-02 山东银河建筑科技有限公司 基于人工智能的城市管理方法及系统
CN116761179A (zh) * 2023-07-14 2023-09-15 青岛科技大学 基于轻量化MS-Net的移动TAS-EH系统安全通信性能智能预测方法
CN116996960A (zh) * 2023-09-26 2023-11-03 江苏星湖科技有限公司 一种电力巡检设备的辅助链路网络接入方法及装置
CN117041133A (zh) * 2023-10-10 2023-11-10 华中科技大学 SRv6网络中流量调度优化方法、系统及介质

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111044285A (zh) * 2019-11-22 2020-04-21 军事科学院系统工程研究院军用标准研究中心 一种复杂情况下机械设备故障诊断方法
CN111179596B (zh) * 2020-01-06 2021-09-21 南京邮电大学 基于组归一化和网格化协同的交通流预测方法
CN111428580A (zh) * 2020-03-04 2020-07-17 威海北洋电气集团股份有限公司 基于深度学习的个体信号识别算法及系统
CN111554089A (zh) * 2020-04-13 2020-08-18 广东省智能制造研究所 一种基于深度学习的交通状态预测方法及装置
CN111950583B (zh) * 2020-06-05 2023-09-26 杭州电子科技大学 一种基于gmm聚类的多尺度识别交通信号标志的方法
CN111800421B (zh) * 2020-07-06 2021-08-24 东北大学 一种基于隐马尔科夫模型的车联网入侵检测系统
CN112750298B (zh) * 2020-12-17 2022-10-28 华路易云科技有限公司 一种基于smdp和drl的货车编队动态资源分配方法
CN113192322B (zh) * 2021-03-19 2022-11-25 东北大学 一种基于云边协同的高速公路交通流量计数方法
CN113380026B (zh) * 2021-05-28 2022-07-22 南京理工大学 基于hmm的高速路交通量预测方法
CN113177536B (zh) * 2021-06-28 2021-09-10 四川九通智路科技有限公司 基于深度残差收缩网络的车辆碰撞检测方法及装置
CN115360719B (zh) * 2022-08-30 2024-04-12 东北大学秦皇岛分校 基于plnn的电力系统短期电压稳定性评估方法
CN115512172B (zh) * 2022-09-30 2023-09-15 赵营鸽 一种用于电阻抗成像技术中多维参数的不确定性量化方法
CN116721551B (zh) * 2023-05-04 2024-05-14 东南大学 基于号牌识别数据的车道级到达曲线历史和实时重构方法
CN117648570B (zh) * 2024-01-30 2024-04-12 吉林大学 一种基于人工智能的心电监测与数据分析系统

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101751782A (zh) * 2009-12-30 2010-06-23 北京大学深圳研究生院 一种基于多源信息融合的十字路口交通事件自动检测系统
CN104408924A (zh) * 2014-12-04 2015-03-11 深圳北航新兴产业技术研究院 一种基于耦合隐马尔可夫模型的城市道路异常交通流检测方法
CN105160866A (zh) * 2015-08-07 2015-12-16 浙江高速信息工程技术有限公司 一种基于深度学习神经网络结构的交通流预测方法
CN106846816A (zh) * 2017-04-12 2017-06-13 山东理工大学 一种基于深度学习的离散化交通状态判别方法
US20190049989A1 (en) * 2017-11-17 2019-02-14 Intel Corporation Identification of audio signals in surrounding sounds and guidance of an autonomous vehicle in response to the same
CN109492814A (zh) * 2018-11-15 2019-03-19 中国科学院深圳先进技术研究院 一种城市交通流量预测方法、系统及电子设备
US10239521B1 (en) * 2018-03-23 2019-03-26 Chongqing Jinkang New Energy Vehicle Co., Ltd. Multi-network-based path generation for vehicle parking

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7610560B2 (en) * 2004-11-16 2009-10-27 Microsoft Corporation Methods for automated and semiautomated composition of visual sequences, flows, and flyovers based on content and context
CN107564313A (zh) * 2017-08-16 2018-01-09 齐鲁交通信息有限公司 高速公路交通拥堵预告系统及预告方法
CN109035779B (zh) * 2018-08-30 2021-01-19 南京邮电大学 基于DenseNet的高速公路交通流预测方法

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101751782A (zh) * 2009-12-30 2010-06-23 北京大学深圳研究生院 一种基于多源信息融合的十字路口交通事件自动检测系统
CN104408924A (zh) * 2014-12-04 2015-03-11 深圳北航新兴产业技术研究院 一种基于耦合隐马尔可夫模型的城市道路异常交通流检测方法
CN105160866A (zh) * 2015-08-07 2015-12-16 浙江高速信息工程技术有限公司 一种基于深度学习神经网络结构的交通流预测方法
CN106846816A (zh) * 2017-04-12 2017-06-13 山东理工大学 一种基于深度学习的离散化交通状态判别方法
US20190049989A1 (en) * 2017-11-17 2019-02-14 Intel Corporation Identification of audio signals in surrounding sounds and guidance of an autonomous vehicle in response to the same
US10239521B1 (en) * 2018-03-23 2019-03-26 Chongqing Jinkang New Energy Vehicle Co., Ltd. Multi-network-based path generation for vehicle parking
CN109492814A (zh) * 2018-11-15 2019-03-19 中国科学院深圳先进技术研究院 一种城市交通流量预测方法、系统及电子设备

Cited By (27)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112308214A (zh) * 2020-11-09 2021-02-02 海南核电有限公司 一种用于冷源致灾物灾害预测的深度学习改进算法
CN112507624A (zh) * 2020-12-15 2021-03-16 交通运输部公路科学研究所 一种城际公路出行方式识别模型构建、识别方法及装置
CN112507624B (zh) * 2020-12-15 2023-11-10 交通运输部公路科学研究所 一种城际公路出行方式识别模型构建、识别方法及装置
CN112836802A (zh) * 2021-02-03 2021-05-25 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) 一种半监督学习方法、岩性预测方法及存储介质
CN113096413A (zh) * 2021-03-23 2021-07-09 北京交通大学 基于高速公路收费数据的省际出行流量估算方法
CN113095381A (zh) * 2021-03-29 2021-07-09 西安交通大学 一种基于改进型dbn的水声目标辨识方法及系统
CN113095381B (zh) * 2021-03-29 2024-04-05 西安交通大学 一种基于改进型dbn的水声目标辨识方法及系统
CN112926273A (zh) * 2021-04-13 2021-06-08 中国人民解放军火箭军工程大学 一种多元退化设备剩余寿命预测方法
CN113240029A (zh) * 2021-05-25 2021-08-10 哈尔滨工程大学 一种gmm-hmm混合模型概率似然得分的svm二次分类方法
CN113537580A (zh) * 2021-06-28 2021-10-22 中科领航智能科技(苏州)有限公司 一种基于自适应图学习的公共交通客流预测方法及系统
CN113537580B (zh) * 2021-06-28 2024-04-09 中科领航智能科技(苏州)有限公司 一种基于自适应图学习的公共交通客流预测方法及系统
CN113612786B (zh) * 2021-08-09 2023-04-07 上海交通大学宁波人工智能研究院 一种车辆总线的入侵检测系统及方法
CN113612786A (zh) * 2021-08-09 2021-11-05 上海交通大学宁波人工智能研究院 一种车辆总线的入侵检测系统及方法
CN113807404B (zh) * 2021-08-24 2024-02-27 中通服咨询设计研究院有限公司 面向多源信息融合的智慧交通网络智能化监控与应急系统
CN113807404A (zh) * 2021-08-24 2021-12-17 中通服咨询设计研究院有限公司 面向多源信息融合的智慧交通网络智能化监控与应急系统
CN114743379A (zh) * 2022-06-13 2022-07-12 广东邦盛北斗科技股份公司 基于北斗的城市大面积路网交通感知方法、系统及云平台
CN115754010B (zh) * 2022-11-10 2024-04-09 四川科力特硬质合金股份有限公司 机械密封环密封端面质量控制方法
CN115754010A (zh) * 2022-11-10 2023-03-07 四川科力特硬质合金股份有限公司 机械密封环密封端面质量控制方法
CN116016065A (zh) * 2023-03-24 2023-04-25 南京大学 基于神经网络的抗串扰高速互联信号接收方法及系统
CN116016065B (zh) * 2023-03-24 2023-08-04 南京大学 基于神经网络的抗串扰高速互联信号接收方法及系统
CN116050672A (zh) * 2023-03-31 2023-05-02 山东银河建筑科技有限公司 基于人工智能的城市管理方法及系统
CN116050672B (zh) * 2023-03-31 2023-06-20 山东银河建筑科技有限公司 基于人工智能的城市管理方法及系统
CN116761179A (zh) * 2023-07-14 2023-09-15 青岛科技大学 基于轻量化MS-Net的移动TAS-EH系统安全通信性能智能预测方法
CN116761179B (zh) * 2023-07-14 2024-04-30 青岛科技大学 基于轻量化MS-Net的移动发射天线选择和能量收集TAS-EH系统安全通信性能智能预测方法
CN116996960A (zh) * 2023-09-26 2023-11-03 江苏星湖科技有限公司 一种电力巡检设备的辅助链路网络接入方法及装置
CN116996960B (zh) * 2023-09-26 2024-01-30 江苏星湖科技有限公司 一种电力巡检设备的辅助链路网络接入方法及装置
CN117041133A (zh) * 2023-10-10 2023-11-10 华中科技大学 SRv6网络中流量调度优化方法、系统及介质

Also Published As

Publication number Publication date
CN110097755A (zh) 2019-08-06
CN110097755B (zh) 2021-08-17
WO2020220439A9 (zh) 2021-02-11

Similar Documents

Publication Publication Date Title
WO2020220439A1 (zh) 基于深度神经网络的高速公路交通流量状态识别方法
CN110120218B (zh) 基于gmm-hmm的高速公路大型车辆识别方法
Kong et al. Weakly labelled audioset tagging with attention neural networks
CN108780523B (zh) 使用本地设备提供的传感器数据和标签的基于云的处理
CN110084151B (zh) 基于非局部网络深度学习的视频异常行为判别方法
CN108921083B (zh) 基于深度学习目标检测的非法流动摊贩识别方法
CN109785618B (zh) 一种基于组合逻辑的短时交通流预测方法
CN106709588B (zh) 预测模型构建方法和设备以及实时预测方法和设备
CN112215487B (zh) 一种基于神经网络模型的车辆行驶风险预测方法
CN103473540B (zh) 智能交通系统车辆轨迹增量式建模与在线异常检测方法
US20190108400A1 (en) Actor-deformation-invariant action proposals
WO2021073524A1 (zh) 一种拥堵交通流溯源分析方法
CN111274886B (zh) 一种基于深度学习的行人闯红灯违法行为分析方法及系统
CN111598325A (zh) 基于层次聚类和分层注意力机制的交通速度预测方法
CN104268546A (zh) 一种基于主题模型的动态场景分类方法
CN113487856B (zh) 基于图卷积网络及注意力机制的交通流组合预测模型
CN105930792A (zh) 一种基于视频局部特征字典的人体动作分类方法
CN104268507A (zh) 一种基于rgb-d图像的手语字母识别方法
CN113256980A (zh) 一种路网状态的确定方法、装置、设备及存储介质
CN111047078A (zh) 交通特征预测方法、系统及存储介质
CN111859010A (zh) 一种基于深度互信息最大化的半监督音频事件识别方法
Wu et al. Traffic pattern modeling, trajectory classification and vehicle tracking within urban intersections
CN112732905A (zh) 一种基于知识图谱的交通事故分析与防控方法及系统
Sun et al. An improved vehicle detection algorithm based on YOLOV3
CN113361491A (zh) 一种无人驾驶汽车对行人的过街意图预测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19927212

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19927212

Country of ref document: EP

Kind code of ref document: A1