A signalized intersection vehicle departure state unsupervised recognition method based on decoupled hidden Markov model

Abstract

The application relates to the technical field of intelligent transportation systems, and discloses a signal intersection vehicle departure state unsupervised identification method based on a decoupling hidden Markov model. In view of the problems of exponential growth of model parameters with state dimension in the existing hybrid state enumeration HMM method and difficult unsupervised training, a three-layer decoupling hidden Markov model composed of a basic state layer, a decoupling state layer and an observation layer is constructed, the departure state is decoupled into two independent dimensions of a basic state and an overflow state, a differentiated observation probability distribution is established, the departure headway and the road travel time correspond to the identification of different types of states, a travel time transformation method is constructed in combination with signal timing information, the recognizability of the overflow state is enhanced, the EM algorithm is used to realize the unsupervised estimation of model parameters, and the Viterbi algorithm and the maximum likelihood criterion are used for state identification. The parameter quantity is reduced by 44%, the state identification accuracy is above 92%, and the application has good robustness and interpretability.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation systems, specifically to an unsupervised method for identifying vehicle departure states at signalized intersections based on a decoupled hidden Markov model. Background Technology

[0002] Signalized intersections are critical nodes in urban road networks and major sources of traffic congestion. Signal control leads to inconsistencies in traffic conditions across time and space, making traffic state identification more challenging. From a signal operation perspective, a key issue is determining whether queued vehicles have dispersed and the efficiency of vehicle departure—the so-called vehicle departure state identification problem. Accurately identifying the departure state of vehicles at signalized intersections—including the initial acceleration process of queued vehicles at the beginning of a green light, the saturated traffic flow after the queue disperses, the free-flow state with no delays, and the overflow state where vehicles experience multiple stops due to queue overflow—is a crucial foundation for achieving real-time traffic monitoring, signal timing optimization, and traffic congestion management.

[0003] In existing technologies, methods based on fixed detector data identify changes in traffic conditions by observing the time intervals between vehicle detections. For example, queue length estimation methods based on shock wave theory identify queue formation, dissipation, and departure shock waves through detector switching events and threshold rules; sequence analysis methods decompose traffic departure flows during green lights into saturated and normal departure flows, using Gaussian process models and Markov chain Monte Carlo techniques to determine flow switching points. However, these methods rely on single-section detector data and have limitations in perceiving traffic conditions beyond the detector location; identifying overflow states when long queues occur remains difficult.

[0004] Emerging floating car trajectory data extends traffic state identification to the spatiotemporal dimensions, but the effectiveness of existing methods largely depends on a high floating car penetration rate (typically above 20%). In the current low-penetration data collection environment, existing studies usually combine historical patterns and real-time data to identify the most likely traffic states, but these methods require a significant amount of prior knowledge of traffic flow dynamics.

[0005] Hidden Markov Models (HMMs) are powerful tools for modeling stochastic processes and have been used to identify vehicle departure states. Previous studies have modeled the time-distance sequence of vehicles leaving the front of a vehicle during a green light as an HMM, encoding 15 hidden states (a mixture of 5 basic departure states and 3 vehicle types). However, the enumeration of mixed states leads to a large number of state variables and parameters, making unsupervised training difficult and requiring manually labeled data. Other studies have extended the HMM framework to integrate travel time data to identify overflow states, introducing covariates to capture the state correlation of the previous cycle and the dynamic state transition matrix, achieving unsupervised self-learning capabilities. However, departure states are only roughly described as overflow, single stop, and free arrival.

[0006] Therefore, the existing technology has the following main shortcomings: (1) In the hidden Markov model method based on mixed state enumeration, the number of hidden states increases exponentially with the dimension of state attributes. Each additional state attribute leads to a doubling of model parameters, making unsupervised training difficult, the convergence speed slow, and difficult to extend to more refined state descriptions.

[0007] (2) Mixed states indicate that the contribution of different categories of observation data to the differential identification of a specific departure state cannot be effectively utilized. For example, the headway of departure is of limited help in distinguishing between overflow state and single stop state, because the headway of departure is mostly at a saturation level in these two states, while the travel time is the key observation to distinguish overflow.

[0008] (3) Existing methods do not make full use of signal timing information to enhance the identification of overflow states. Overflow vehicles experience additional red light delays, and their travel time differs systematically from that of normal vehicles. However, the original travel time distribution may overlap, making it difficult to distinguish them directly.

[0009] (4) Under the condition of limited and noisy actual observation data, the existing model lacks sufficient robustness, especially when the missing rate of travel time observations is high, the performance drops significantly. Summary of the Invention

[0010] This invention provides an unsupervised identification method for vehicle departure states at signalized intersections based on a decoupled Hidden Markov Model. By decoupling the departure state into a basic state and a decoupled state, and utilizing different types of observation data respectively, the method achieves simplified model parameters and fully self-supervised state identification.

[0011] To achieve the above objectives, an unsupervised method for identifying vehicle departure states at signalized intersections based on a decoupled Hidden Markov Model is provided, comprising the following steps: S1: Obtain vehicle passage records collected by upstream and downstream LPR sensors at adjacent signalized intersections, extract the time distance between vehicle departures and the travel time of the road segment, and preprocess abnormal travel times.

[0012] LPR sensors are deployed at the stop lines upstream and downstream of adjacent signalized intersections. When a vehicle crosses the stop line, the LPR sensors simultaneously record the vehicle's passage time, license plate number, lane information, and signal cycle information. By matching the same license plate upstream and downstream, a vehicle-level observation sequence containing the time distance before the vehicle leaves the lane and the travel time along the road segment can be generated.

[0013] Based on the order in which vehicles pass the downstream stop line within the same green light cycle, the time distance between the vehicles leaving the front is extracted. Let the time when the nth vehicle passes the downstream stop line and is identified in the i-th signal cycle be denoted as . The green light for this cycle begins at [time]. The total number of vehicles leaving during the period was The distance at which the vehicle leaves the front of the vehicle is defined as: ; in, This represents the time distance between the nth vehicle and the front of the vehicle within the i-th period; for the first vehicle within the period, This indicator reflects the start-up response time from when the green light turns on to when a vehicle crosses the stop line. It is primarily used to characterize the time characteristics of continuous vehicle departure during start-up, saturation, and free-flow conditions.

[0014] By matching license plates in the LPR records at upstream and downstream intersections, the travel time of the same vehicle from exiting the upstream stop line to exiting the downstream stop line is calculated. Let... The moment when the vehicle crosses the upstream stop line and is identified. If the vehicle crosses the downstream stop line and is identified at the moment, then: ; in, This indicates the travel time of a vehicle between adjacent intersections. Compared to the time it takes to leave the front of the vehicle, the travel time is more likely to reflect whether the vehicle has experienced additional red light delays, and is therefore mainly used for identifying overflow conditions.

[0015] Outlier filtering is performed on the original travel times. For abnormal travel times caused by non-signal control factors such as bus stops, temporary parking, vehicles entering or leaving the route midway, and license plate recognition errors, a first-in-first-out (FIFO) filter based on the Longest Increasing Subsequence (LIS) algorithm is used to remove them; at the same time, minimum and maximum travel time thresholds are used to further filter out abnormal samples. For vehicles that cannot be matched simultaneously upstream and downstream, the time distance between their departure from the front of the vehicle is retained, and their travel time is marked as a missing observation.

[0016] S2: Based on the mechanism of vehicles leaving during the green light period, a three-layer decoupled hidden Markov model is constructed, consisting of a basic state layer, a decoupled state layer, and an observation layer.

[0017] Define the basic state layer. The vehicle departure process during the green light period is divided into three basic states: start-up, saturation, and free-flow, denoted as start-up, saturation, and free-flow, respectively. Let the basic departure state of the nth vehicle be... Then the basic state sequence A Markov chain is constructed to describe the stochastic evolution of a vehicle's departure state as it departs from its departure sequence.

[0018] Define a decoupling state layer. Considering that the overflow state may occur simultaneously with the startup state or the saturation state, this invention no longer treats combinations such as "startup + overflow" or "saturation + non-overflow" as independent mixed states, but instead introduces a separate decoupling state. This indicates whether the vehicle is in an overflow state. Among them, Indicates an overflow state. This indicates a non-overflow state. The decoupled state primarily affects the travel time distribution of road segments, without directly altering the left-to-right transition structure of the basic state.

[0019] Define the observation layer. Each vehicle corresponds to two observation variables: the time distance from the front of the vehicle. and travel time Basic state The main factor determining the distance when the vehicle leaves the front is... Conditional probability distribution; decoupling state The main factor determining the trip duration The conditional probability distribution.

[0020] Based on the above definitions of state and observation, the joint probability distribution of the decoupled Hidden Markov Model can be expressed as: ; in, Let be the initial basic state probabilities. Let be the transition probability between the basic states of adjacent vehicles. Let be the probability that the vehicle is in an overflow or non-overflow state given the basic state conditions. and These are the conditional probability distributions of travel time and headway, respectively. Let represent the initial probability of the basic state k. This represents the transition probability from basic state j to k. This represents the probability of the decoupled state l occurring given the basic state k. and Let represent the mean and standard deviation of the Gaussian distribution when the vehicle leaves the front of the car under the basic state k. and Let represent the mean and standard deviation of the Gaussian distribution of the travel time of the road segment under the decoupled state l, respectively.

[0021] To ensure that the state evolution process conforms to the vehicle departure mechanism at signalized intersections, the basic state layer adopts a left-to-right transition constraint. Specifically, the starting state can remain in itself or transition to a saturated state, the saturated state can remain in itself or transition to a free-flow state, and the free-flow state can remain in itself; upon entering the next signal cycle, the state sequence restarts from the starting state. These constraints can be implemented by defining an upper diagonal state transition matrix, which avoids physically meaningless random state jumps and enhances the stability of the unsupervised training process.

[0022] S3: Establish observation probability distributions based on the time distance from the vehicle head and the travel time of the road segment, so that different observation data correspond to the identification tasks of different types of departure states.

[0023] The distance between the vehicle and the front of the vehicle is modeled using a Gaussian distribution. For the basic state... ,have: ; in, and denoted as the mean and standard deviation of the headway distance when a vehicle leaves the front of the vehicle under the k-th basic state. The starting state typically has a large and decreasing headway distance, the saturation state has a relatively stable headway distance, and the headway distance in the free-flow state is more affected by the randomness of vehicle arrival.

[0024] The vehicle travel time on road segments is modeled using a decoupled state distribution. Given the basic state... and decoupling state When the travel time probability term is: ; in, This represents the probability that the vehicle is in the l-th type of decoupled state given the base state k; and denoted as the mean and standard deviation of the travel time distribution under the l-th type of decoupling state, respectively.

[0025] Based on the properties of summation and product, the marginal conditional distribution of travel time under given basic state conditions can be obtained: ; This distribution resembles a mixture Gaussian distribution, but the Gaussian components are shared by different fundamental states, and the fundamental states are identified only by their coefficients. Choose different mixing ratios. Compared to directly enumerating the mixing states, this approach can significantly reduce the number of model parameters while maintaining the ability to characterize complex distributions, making it more suitable for unsupervised training under conditions of limited samples and noisy observations.

[0026] S4: Combine signal timing information to construct a travel time transformation method, and enhance the difference between overflow state and non-overflow state by expecting to control delay.

[0027] Since the original travel time may change continuously with the vehicle departure sequence, the travel time distribution of overflowing vehicles and non-overflowing vehicles may overlap, and directly using the original travel time to identify the overflow status is prone to misjudgment. Therefore, this invention introduces a travel time transformation based on desired control delay to enhance the travel time deviation caused by overflowing vehicles experiencing additional red light delays.

[0028] Let the red light duration of the i-th signal cycle be... The green light duration is The time when a vehicle departs from the station relative to the start of the green light within this cycle is Therefore, the expected delay is calculated based on a linear decreasing relationship within the green light interval, and the original travel time is transformed accordingly: , ; in, This indicates a desire to control delays. This represents the transformed travel time. This transformation is not intended to accurately estimate the actual control delay for each vehicle, but rather to amplify the separability of travel time characteristics between overflowing and non-overflowing vehicles.

[0029] For non-overflow vehicles, it is expected that the control delay can absorb the normal signal control effects they experience in the current cycle, making the transformed travel time distribution more concentrated. For overflow vehicles, due to the residual queues and additional red light delays from the previous cycle, a significant additional delay deviation remains after the transformation. Therefore, travel time transformation can improve the identifiability of overflow status.

[0030] S5: The Expectation Maximization (EM) algorithm is used to perform unsupervised estimation of the model parameters, and the posterior probability of the hidden state is calculated by forward-backward variable recursion.

[0031] Let the set of model parameters to be estimated be... Given the travel time observation sequence T and the headway observation sequence H, the model parameters are estimated by maximizing the likelihood function of the observed data:

[0032] Since both the basic state sequence S and the decoupled state sequence M are latent variables, it is difficult to directly obtain a closed-form solution for the maximum likelihood estimation. This invention employs the expectation-maximization algorithm for iterative solution, where the E-step calculates the posterior probabilities of the latent states, and the M-step updates the model parameters under the condition that the posterior probabilities are fixed.

[0033] In the E-step, the posterior probabilities of the vehicle being in the basic state, the decoupled state, and the transition between adjacent basic states are defined as follows:

[0034] in, Let represent the posterior probability that the nth vehicle is in the basic state k; This represents the joint posterior probability that the nth vehicle is simultaneously in both the basic state k and the decoupled state l. Let represent the posterior probability of the state transition when the (n-1)th car is in base state j and the nth car is in base state k. Initial parameters It can be determined by Gaussian mixture models or by statistical analysis of a small number of field observations.

[0035] In the M-step, the initial state probability, the basic state transition probability, and the decoupled state probability are updated based on the aforementioned posterior probabilities:

[0036] The new parameters obtained in each M-step are used as inputs for the next E-step, until the change in model parameters or the increment of the likelihood function of the observed data satisfies the preset convergence condition.

[0037] To improve the efficiency of posterior probability calculation, a forward variable is introduced. and backward variables :

[0038] By using forward-backward recursion, we can avoid enumerating all possible hidden state sequences, thus achieving efficient unsupervised training under multi-cycle and multi-vehicle sequence conditions.

[0039] S6: After completing the parameter estimation, the Viterbi algorithm and the maximum likelihood criterion are used to identify the basic vehicle departure state and the decoupling overflow state, respectively.

[0040] After the model parameters are estimated, the Viterbi algorithm is first used to identify the most likely sequence of basic states. The Viterbi algorithm expands the state nodes of each vehicle in the departure order and searches for the path with the maximum posterior probability in the state transition constraint graph. ; in, The most likely sequence of basic states is represented, with the vehicle identified as either in a starting state, a saturated state, or a free-flowing state.

[0041] After determining the basic states, the decoupling states are identified using the maximum likelihood criterion. For the nth vehicle, its overflow state determination can be expressed as: ; like Then the vehicle is identified as an overflow vehicle; if If the condition is met, the vehicle is identified as a non-overflow vehicle. This allows for the generation of combined interpretations for each vehicle, such as "startup + overflow," "saturation + non-overflow," or "free flow."

[0042] When vehicle travel time for a road segment is missing, the model can still infer the basic departure state using distance observations from the vehicle's headway and basic state transition relationships. In this case, the likelihood of the missing travel time for the road segment is 1, indicating that the missing data does not affect the overall likelihood. When travel time observations are available, they are further used to enhance overflow state identification. Therefore, this invention has a certain tolerance for actual data loss situations such as LPR matching failure, lack of upstream detectors in some lanes, and vehicles entering or leaving midway.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows: Compared with Hidden Markov Models (HMMs) based on mixed state enumeration, this invention uses decoupled state representation, reducing the number of model parameters from 50 to 28, and the training convergence time is about half that of the former, effectively solving the problems of exponential parameter growth and unsupervised training difficulties caused by mixed state enumeration. Compared with the decoupled model that does not perform travel time transformation, the present invention significantly improves the state identification accuracy on two road segments by using travel time transformation based on signal timing. In particular, the identification accuracy is improved by 7.3 and 8.0 percentage points for the two states of "start-up + overflow" and "saturation + overflow", respectively, which verifies the effectiveness of travel time transformation in enhancing the identifiability of overflow state. The decoupled state representation of this invention can clearly present the physical state evolution pattern of "start-up → saturation → free flow". Posterior probability analysis shows that the probability of the start-up state tends to zero after the 5th vehicle, the saturation probability first increases and then decreases, and the free flow probability gradually increases, which has a clear traffic physics meaning and interpretability, while the traditional mixed-state HMM model shows state identification ambiguity and random transition phenomena. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the overall process of vehicle departure status identification at signalized intersections based on a decoupled hidden Markov model, as described in this invention. Figure 2 This is a schematic diagram of the LPR sensor data acquisition scenario of the present invention; Figure 3 This is a schematic diagram of the decoupled hidden Markov model structure and state transition constraints of the present invention; Figure 4 This is a schematic diagram of the travel time transformation based on the desired control delay of the present invention; Figure 5 This is a comparison chart of the departure state identification results of different models in this invention; Figure 6 This is a comparison diagram of the posterior probability evolution of the departure state sequence in this invention; Figure 7 This is a diagram showing the expected number of vehicles in each departure state according to the present invention; Figure 8 This is a comparison chart of the state identification accuracy of each model under different travel time missing rates in this invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer and more complete, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention in any way. Example

[0046] This embodiment selects two adjacent signalized intersections on a main road in a city as the test site. The road section between the upstream and downstream intersections is 805 meters long, and LPR sensors (electronic traffic enforcement devices) are installed at both the upstream and downstream intersections. The analysis period is selected from 16:30 to 18:30 on weekday evenings, during which traffic flow is relatively large and can cover four departure states: starting, saturation, free flow, and overflow.

[0047] The data collection process is as follows: LPR sensors are installed at the stop lines of each approach lane at the upstream and downstream intersections. When a vehicle passes the stop line, the LPR sensor simultaneously records the following information: the time of vehicle passage (accurate to 0.1 seconds), license plate number, lane number, signal cycle number, and current signal light color status.

[0048] Based on the order in which vehicles passed the downstream stop line, the vehicle departure records within the same green light cycle are sorted by time. Let the time when the nth vehicle in the i-th signal cycle passes the downstream stop line be . The green light for this cycle begins at [time]. The total number of vehicles leaving during the period was The distance between the vehicle and the front of the vehicle is calculated using the following formula: ;

[0049] By matching license plates in the LPR records at upstream and downstream intersections, the travel time of the same vehicle from exiting the upstream stop line to exiting the downstream stop line is calculated. Let... The moment when the vehicle crosses the upstream stop line and is identified. If the vehicle crosses the downstream stop line and is identified at the moment, then:

[0050] Outlier filtering is performed on the original travel times. For abnormal travel times caused by non-signal control factors such as bus stops, temporary parking, vehicles entering or leaving the route midway, and license plate recognition errors, a first-in-first-out (FIFO) filter based on the Longest Increasing Subsequence (LIS) algorithm is used to remove them; at the same time, minimum and maximum travel time thresholds are used to further filter out abnormal samples. For vehicles that cannot be matched simultaneously upstream and downstream, the time distance between their departure from the front of the vehicle is retained, and their travel time is marked as a missing observation.

[0051] Step 2: Construction of a three-layer decoupled hidden Markov model Based on the mechanism of vehicles leaving signalized intersections during green light periods, a three-layer decoupled hidden Markov model is constructed.

[0052] Define the basic state layer. The vehicle departure process during the green light period is divided into three basic states: starting state, saturation state, and free-flow state. Let the basic departure state of the nth vehicle be... Then the basic state sequence This forms a Markov chain.

[0053] Define a decoupling state layer. Introduce decoupling states separately. This indicates whether the vehicle is in an overflow state. Among them, Indicates an overflow state. This indicates a non-overflow state.

[0054] Define the observation layer. Each vehicle corresponds to two observation variables: the time distance from the front of the vehicle. and travel time Basic state When deciding to drive away from the front of the car Conditional probability distribution; decoupling state Decide on travel time The conditional probability distribution.

[0055] The joint probability distribution of the decoupled Hidden Markov Model is:

[0056] To ensure that the state evolution process conforms to the vehicle departure mechanism of a signalized intersection, the basic state layer adopts a left-to-right transition constraint. The starting state can remain in its current state or transition to a saturated state; the saturated state can remain in its current state or transition to a free-flowing state; the free-flowing state can remain in its current state. Upon entering the next signal cycle, the state sequence restarts from the starting state. These constraints are implemented by defining an upper diagonal state transition matrix: .

[0057] Step 3: Parameterization of the Observation Probability Distribution The distance between the vehicle and the front of the vehicle is modeled using a Gaussian distribution. For the basic state... ,have:

[0058] The vehicle travel time on road segments is modeled using a decoupled state distribution. Given the basic state... and decoupling state When the travel time probability term is:

[0059] Given the basic state k, the marginal conditional distribution of travel time is:

[0060] Step 4: Travel-time transformation based on signal timing Let the red light duration of the i-th signal cycle be... The green light duration is The time when a vehicle departs from the station relative to the start of the green light within this cycle is Therefore, the expected delay is calculated based on a linear decreasing relationship within the green light interval, and the original travel time is transformed accordingly:

[0061] Step 5: Unsupervised estimation of model parameters based on the EM algorithm An iterative solution is obtained using the expectation-maximization algorithm. In the E-step, forward variables are utilized. and backward variables Calculate the posterior probability recursively:

[0062] In the M-step, the model parameters are updated based on the posterior probabilities described above:

[0063] Iteratively execute the E-step and M-step until the change in model parameters or the increment of the likelihood function of the observed data satisfies the preset convergence condition.

[0064] Step Six: Vehicle Departure Status Identification After the model parameters are estimated, the Viterbi algorithm is used to solve for the most likely sequence of basic states:

[0065] After determining the basic state, the decoupling state is determined based on the maximum likelihood criterion:

[0066] like Then the vehicle is identified as an overflow vehicle; if If so, the vehicle is identified as a non-overflow vehicle.

[0067] When the travel time of a vehicle segment is missing, the model can still infer the basic departure state by using the distance observation of the vehicle leaving the front and the basic state transition relationship. The likelihood value of the missing travel time segment is 1.

[0068] Example 2 To verify the effectiveness of the method of this invention, a field experiment was conducted on two urban arterial road sections. Road section A is 805m long, and road section B is 359m long. The analysis period was weekday evening rush hour from 16:30 to 18:30. The performance of the method of this invention (T-DHMM) was compared with the following two comparative methods: a traditional mixed-state HMM (which enumerates combinations of start-up, saturation, free flow, and overflow states as independent implicit states) and a decoupled DHMM that does not perform travel-time transformation.

[0069] Evaluation metrics include maximum log-likelihood (MLL), state identification accuracy (SIA), and training time. State identification accuracy is defined as:

[0070] in, The total number of vehicles available for verification is determined by manual data collection from video surveillance of actual vehicle departures.

[0071] The experimental results are shown in Table 1: Table 1 Comparison of field experimental performance of different models

[0072] As shown in Table 1, T-DHMM achieved the highest maximum log-likelihood and state identification accuracy on both road segments, indicating that the decoupled state representation can better interpret LPR observation data. Compared with mixed-state HMM, T-DHMM has a simpler model structure, significantly shorter training time, and a more stable unsupervised parameter estimation process.

[0073] The accuracy rates for identifying each departure status are shown in Table 2: Table 2 Comparison of recognition accuracy of T-DHMM and DHMM for various departure states

[0075] As shown in Table 2, the T-DHMM improved the recognition accuracy of the "start-up + overflow" and "saturation + overflow" states by 7.3 and 8.0 percentage points, respectively, indicating that the travel-time transformation based on signal timing mainly improved the recognition effect of overflow-related states.

[0076] The results in Tables 1 and 2 show that T-DHMM can exhibit a clear "start-up → saturation → free-flow" state evolution pattern: the probability of the start-up state gradually decreases after the first few vehicles, the probability of the saturation state first increases and then decreases, and the probability of the free-flow state gradually increases as the departure sequence extends; moreover, T-DHMM maintains an identification accuracy of over 80% in all scenarios; under the same missing rate, T-DHMM maintains a higher identification accuracy than DHMM and HMM overall.

[0077] Example 3 This embodiment further constructs a simulation scenario based on the SUMO micro-simulation platform to examine the applicability of the model under different traffic volumes, different proportions of heavy vehicles, and different signal coordination conditions. In the simulation experiment, the travel time missing rate (MR) was set to 0.1-0.5, and the simulation parameters used were the default SUMO parameters.

[0078] Table 3 shows the accuracy of T-DHMM state identification under different simulation scenarios: Table 3. T-DHMM State Identification Accuracy under Different Simulation Scenarios

[0079] As shown in Table 3, the model can achieve a high identification accuracy under conditions of moderate traffic volume and low missing rate. When the traffic volume is low, there are fewer overflow samples, and the model's learning of the overflow state distribution is relatively insufficient. When the proportion of heavy vehicles increases, the larger headway and potential mobility bottleneck effects increase the uncertainty of basic state identification. When the travel time missing rate increases, the key observations required for overflow state identification decrease, and the model performance declines.

[0080] Based on the combined results of field experiments and simulation experiments, this invention can achieve high-precision unsupervised identification of vehicle departure status while maintaining a simple model structure, and demonstrates good robustness and adaptability in various real traffic scenarios.

[0081] Example 4 Based on the state identification results, the expected number of vehicles in different departure states within each signal cycle can be calculated: ;

[0082] in, This represents the expected number of vehicles under the combination of the k-th basic state and the l-th decoupled state. This metric can be used to evaluate the queue dissipation process and the effectiveness of signal control operations at signalized intersections.

[0083] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An unsupervised method for identifying vehicle departure states at signalized intersections based on a decoupled hidden Markov model, characterized in that, Includes the following steps: Acquire vehicle passage records collected by license plate recognition sensors upstream and downstream of signalized intersections, and extract vehicle departure time sequence and road segment travel time sequence; A three-layer decoupled hidden Markov model is constructed, comprising a basic state layer, a decoupled state layer, and an observation layer. The basic state layer includes a starting state, a saturated state, and a free-flow state, which are used to describe the Markov chain of the vehicle departure state as it evolves with the departure sequence number. The decoupled state layer includes overflow states and non-overflow states, which are used to describe whether the vehicle experiences additional red light delays. The observation layer includes departure headway observations and road segment travel time observations. Establish a differentiated observation probability distribution so that the basic state layer determines the conditional probability distribution of the time distance from the front of the vehicle, and the decoupled state layer determines the conditional probability distribution of the travel time of the road segment. The travel time of the road segment is transformed based on signal timing information to enhance the distinguishability between overflow and non-overflow states. The parameters of the three-layer decoupled hidden Markov model are estimated using the expectation-maximization algorithm, and the posterior probability of the hidden state is calculated by forward-backward variable recursion. After parameter estimation, the Viterbi algorithm is used to identify the basic departure state of the vehicle, and the overflow state of the vehicle is determined based on the maximum likelihood criterion.

2. The method according to claim 1, characterized in that, When acquiring vehicle traffic records collected by upstream and downstream vehicle identification sensors at a signalized intersection and extracting the vehicle departure time sequence and road segment travel time sequence, the vehicle departure time sequence is defined as the time interval between two consecutive vehicles leaving the intersection and passing the downstream stop line within the same green light cycle. The departure time sequence of the first vehicle within the cycle is the time difference between the start of the green light and the moment the vehicle passes the stop line. The road segment travel time is obtained by matching the identification records of the same vehicle by upstream and downstream vehicle identification sensors, and is the time difference between the moment the same vehicle passes the upstream stop line and the moment it passes the downstream stop line. Outlier filtering is performed on the road segment travel time. A first-in-first-out (FIFO) filter based on the longest increasing subsequence algorithm is used to remove travel time samples that do not meet the FIFO constraint. The minimum and maximum travel time thresholds are combined to filter out abnormal samples. The travel time of vehicles that cannot be matched upstream and downstream is marked as a missing observation.

3. The method according to claim 1, characterized in that, The joint probability distribution of the three-layer decoupled hidden Markov model in step S2 is expressed as follows: ； in, For the basic state sequence, These represent the startup state, saturation state, and free-flow state, respectively. For the decoupled state sequence, These represent the non-overflow state and the overflow state, respectively; The distance between the tanker truck and the locomotive when the tanker truck leaves the locomotive; This is a sequence of travel time observations for road segments; Let be the initial basic state probabilities. The transition probability between the basic states of adjacent vehicles; Given the basic state conditions, this represents the probability that the vehicle belongs to either an overflow or non-overflow state. This represents the conditional probability distribution of travel time in a decoupled state. Conditional probability distribution of the headway distance under basic conditions.

4. The method according to claim 1, characterized in that, When establishing the differential observation probability distribution, the conditional probability distribution of the time distance from the front of the vehicle is modeled using a Gaussian distribution as follows: ； in, and These are the mean and standard deviation of the distance between the vehicle and the front of the vehicle when the vehicle leaves the front of the vehicle under the k-th basic state, respectively; The conditional probability distribution of the travel time for the aforementioned road segment is modeled using a decoupled state Gaussian distribution as follows: ； in, and These are the mean and standard deviation of the travel time of the road segment under the l-th decoupling state, respectively; Given the basic state k, the marginal conditional distribution of travel time is: ； in, This represents the probability that the vehicle is in the l-th type of decoupled state given a base state k.

5. The method according to claim 1, characterized in that, When changing the travel time of the road segment based on signal timing information: Let the red light duration of the i-th signal cycle be... The green light duration is The time when a vehicle departs from the station relative to the start of the green light within this cycle is The expected control delay is calculated based on a linear decreasing relationship within the green light interval: ； Transform the original route travel time: ; in, This indicates a desire to control delays. Indicates the original route travel time. The transformed travel time is indicated; the transformation is used to amplify the travel time deviation caused by overflow vehicles experiencing additional red light delays, while making the travel time distribution of non-overflow vehicles more concentrated.

6. The method according to claim 1, characterized in that, When using the expectation-maximization algorithm to perform unsupervised estimation of model parameters, in the E-step, based on the current model parameters, the forward variables are used. and backward variables Calculate the posterior probability recursively: ； Let represent the posterior probability of the nth vehicle being in basic state k, the joint posterior probability of the nth vehicle being in both basic state k and decoupled state l, and the transition posterior probability of the (n-1)th vehicle being in basic state j and the nth vehicle being in basic state k, respectively. In the M-step, the initial state probability is updated in a closed loop based on the posterior probability. Basic state transition probability Decoupling state probability Parameters of headway distribution and travel time distribution parameters ; Iteratively execute the E-step and M-step until the change in model parameters or the increment of the likelihood function of the observed data satisfies the preset convergence condition.

7. The method according to claim 1, characterized in that, When identifying the vehicle's departure status after completing parameter estimation The Viterbi algorithm is used to solve for the basic state sequence with maximum a posteriori probability. ； Each vehicle is identified as either in a starting state, a saturated state, or a free-flow state; after determining the basic states, the decoupling state is determined based on the maximum likelihood criterion. ； like Then the vehicle is identified as an overflow vehicle, if The vehicle is then identified as a non-overflow vehicle; the combined state of each vehicle is output, including "start + overflow", "start + non-overflow", "saturation + overflow", "saturation + non-overflow" or "free flow".

8. The method according to claim 1, characterized in that, When the travel time observations of the vehicle on the aforementioned road segment are missing, the three-layer decoupled hidden Markov model still uses the distance observations of leaving the vehicle head and the basic state transition relationship to infer the basic departure state. The likelihood of the missing road segment travel time is set to 1, which does not affect the overall likelihood. When the road segment travel time observations are available, they are further used to enhance the identification of overflow states.

9. The method according to claim 1, characterized in that, The basic state layer adopts a left-to-right transition constraint, that is, the starting state can maintain itself or transition to the saturated state, the saturated state can maintain itself or transition to the free flow state, and the free flow state can only maintain itself. After entering the next signal cycle, the basic state sequence starts again from the starting state; the transition constraint is implemented by defining the upper diagonal state transition matrix.

10. The method according to claim 1, characterized in that, It also includes calculating the expected number of vehicles in different departure states within each signal cycle based on the model parameter estimation results and state identification results: ； in This represents the expected number of vehicles under the combination of the k-th basic state and the l-th decoupled state; the expected number of vehicles is used to evaluate the queue dissipation process and signal control operation effect at the signalized intersection.

Images