A smart traffic supervision system and method based on an internet of things

Abstract

This invention discloses an IoT-based intelligent traffic monitoring system and method, relating to the field of intelligent transportation technology. It collects vehicle trajectory data, navigation path data, traffic flow data, and road environment data through IoT sensing devices deployed in the road environment. Based on comparative analysis of vehicle trajectories and navigation paths, combined with real-time road conditions and passenger behavior characteristics, multiple behavioral intent parameters are calculated. Trajectory features and behavioral intent parameters are fused to construct a trajectory behavior vector, which is then input into a deep trajectory clustering model to complete initial clustering. A context-aware weight matrix is ​​generated based on road type, time period, and vehicle attributes, dynamically adjusting the clustering boundaries to construct flexible clustering regions. The warning results are compared with actual enforcement data, and the weights of behavioral intent parameters and the clustering model structure are dynamically adjusted based on feedback labels to improve the system's recognition accuracy and stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, specifically to an intelligent transportation monitoring system and method based on the Internet of Things. Background Technology

[0002] Intelligent traffic management refers to the use of advanced information technology, the Internet of Things, big data, artificial intelligence, and other means to monitor, analyze, and manage traffic operation status, traffic facilities, and traffic participants in real time, achieving intelligent, refined, and efficient supervision of traffic operations. Its purpose is to improve traffic safety, alleviate traffic congestion, optimize resource allocation, and promote the development of urban transportation systems towards intelligence.

[0003] Existing deep trajectory clustering models suffer from a "potential multi-intention overlap" problem when identifying abnormal vehicle paths. This means the model relies on historical trajectory distributions for clustering decisions and fails to recognize reasonable deviations caused by navigation to avoid congestion, passenger requests, or temporary tasks. This leads to the misclassification of normal but rare trajectories as abnormal behavior. This problem is particularly severe in scenarios with diverse travel patterns and complex urban road networks, not only reducing model accuracy but also potentially causing misjudgments by law enforcement, impacting the fairness and credibility of traffic supervision. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent traffic monitoring system and method based on the Internet of Things to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a smart traffic monitoring method based on the Internet of Things, comprising: Collect multi-source traffic operation information, including vehicle trajectory data, navigation route data, traffic flow data, and road environment data; Based on the comparative analysis of vehicle trajectory data and navigation path data, combined with real-time traffic conditions and passenger behavior characteristics, behavioral intent parameters used to characterize the rationality of vehicle driving behavior are calculated. The original trajectory features and behavioral intent parameters are jointly embedded into the input vector of the deep trajectory clustering model to construct the trajectory behavior vector space; A deep trajectory clustering model is used to perform clustering analysis on trajectory behavior vectors, and the clustering boundaries are dynamically adjusted based on road type, time period and vehicle attribute context elements to construct elastic clustering regions for identifying potential abnormal trajectories. The warning results are compared with the subsequent actual law enforcement results, and the weights of the behavioral intent parameters and the structure of the deep trajectory clustering model are updated online.

[0006] Preferably, the behavioral intent parameters include navigation deviation rate, congestion avoidance factor, passenger intervention index, time rarity score, and behavioral consistency score.

[0007] Preferably, the original trajectory features and behavioral intent parameters are jointly embedded into the input vector of the deep trajectory clustering model to construct a trajectory behavior vector space, specifically including: Time series modeling is performed on the collected vehicle trajectory data to extract spatiotemporal behavioral features, including trajectory point sequences, speed changes, acceleration patterns, turning frequencies, and stopping behaviors. Recurrent neural networks or autoencoders are then used to vectorize the spatiotemporal behavioral features to generate trajectory base coding vectors. The calculated behavioral intent parameters are normalized and used as additional semantic dimensions, and then fused and concatenated with the trajectory base encoding vector at the feature layer to form a multi-dimensional joint feature representation. Multi-dimensional joint feature representations are embedded as input into a pre-defined deep trajectory clustering model, and the deep trajectory clustering model is used to construct a trajectory behavior vector space with behavioral intent semantic constraints in a high-dimensional space.

[0008] Preferably, a deep trajectory clustering model is used to perform clustering analysis on trajectory behavior vectors, and the clustering boundaries are dynamically adjusted based on road type, time period, and vehicle attribute contextual elements, specifically: The deep trajectory clustering model is used to receive trajectory behavior vectors as input, and the trajectory behavior vectors include the spatiotemporal trajectory encoding features of the vehicle and normalized behavioral intent parameters. Through a deep embedding mechanism, multiple input vectors are mapped to a high-dimensional feature space, and the samples in this space are initially clustered to form multiple basic trajectory clusters. Collect contextual element information related to the trajectory, including road type information, time period information, and vehicle attribute information; construct a context-aware weight matrix based on different context combinations, and apply this matrix to the boundary representation of the basic trajectory cluster; Based on the context-aware weight matrix adjustment results, elastic clustering regions are generated. If a trajectory behavior vector is outside the boundary of the main cluster but still within the elastic extension zone, it is not considered an abnormal trajectory. The abnormal trajectory alarm mechanism is triggered only when the trajectory sample exceeds the boundary of the elastic zone or when multiple behavioral intent parameters simultaneously show a high risk of abnormality.

[0009] Preferably, based on the adjustment results of the context-aware weight matrix, elastic clustering regions are generated, specifically including: For each trajectory cluster generated by the deep trajectory clustering model, extract its boundary feature information, including the cluster center vector, boundary radius, intra-cluster sample density distribution, and deviation threshold of feature dimensions; Based on the pre-generated context-aware weight matrix, weight adjustment factors are assigned to the key boundary feature dimensions in different clusters. The weight matrix is ​​dynamically generated according to the road type to which the vehicle belongs, the current time period, and the vehicle attribute category. Based on the mapping relationship between the weight matrix and the boundary features, the elasticity coefficient of each cluster in each dimension is calculated to represent the acceptable deviation range of that dimension in the current context. By using elasticity coefficients to non-uniformly expand the basic cluster boundaries, elastic cluster boundaries with anisotropic variable radii and density thresholds are constructed, forming the final elastic cluster regions.

[0010] Preferably, the warning results are compared with the subsequent actual law enforcement results, and the weights of the behavioral intent parameters and the structure of the deep trajectory clustering model are updated online, specifically including: After identifying abnormal trajectories and generating early warnings based on trajectory behavior vectors and elastic clustering regions, actual law enforcement results data related to the early warning event are collected. Enforcement results data include manual review results, on-site enforcement records, post-event behavior verification information, or user complaint feedback; the warning conclusions and enforcement conclusions of each trajectory sample are compared with the labels to form "hit / false alarm" pair; Based on the scoring results of trajectory samples across multiple behavioral intent parameters and the corresponding hit / false alarm pairs, the weight coefficients of each behavioral intent parameter are adjusted. If a parameter is highly sensitive in false positive samples, its weight is reduced; conversely, if it is significantly discriminative in correct samples, its weight is increased. By combining the accumulated feedback label samples, they are re-embedded into the trajectory behavior vector space, and the embedding layer, cluster center vector, and cluster boundary parameters of the deep trajectory clustering model are adjusted.

[0011] This invention also provides an Internet of Things-based intelligent traffic monitoring system, comprising: The data acquisition module collects multi-source traffic operation information, including vehicle trajectory data, navigation path data, traffic flow data, and road environment data. The behavioral intent parameter calculation module calculates behavioral intent parameters to characterize the rationality of vehicle driving behavior based on the comparative analysis of vehicle trajectory data and navigation path data, combined with real-time traffic conditions and passenger behavior characteristics. The trajectory behavior vector space construction module jointly embeds the original trajectory features and behavior intent parameters into the input vector of the deep trajectory clustering model to construct the trajectory behavior vector space; The analysis and adjustment module uses a deep trajectory clustering model to perform clustering analysis on trajectory behavior vectors, and dynamically adjusts the clustering boundaries based on road type, time period and vehicle attribute context elements to construct elastic clustering regions for identifying potential abnormal trajectories. The update module compares the early warning results with the subsequent actual law enforcement results and updates the weights of the behavioral intent parameters and the structure of the deep trajectory clustering model online.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention introduces a closed-loop adaptive optimization mechanism based on actual law enforcement feedback, streamlining the entire process from abnormal trajectory identification and behavioral intent parameter modeling to early warning result verification, dynamic weight adjustment, and cluster model fine-tuning. Compared to traditional static rules or fixed-weight models, this invention's system possesses self-learning, self-correction, and adaptive capabilities, continuously optimizing the accuracy of identifying atypical but reasonable trajectories, significantly reducing false alarm rates, and effectively improving the intelligence level and reliability of traffic monitoring systems.

[0013] 2. The context-aware clustering boundary dynamic adjustment method and elastic clustering region construction strategy proposed in this invention enable the clustering model to automatically expand its recognition tolerance according to different road types, time periods, and vehicle attributes, thereby enhancing the model's generalization ability and fault tolerance performance in multiple scenarios. The overall technical solution not only improves the flexibility and practicality of the trajectory recognition model, but also enhances the system's ability to be implemented and its scalability in real traffic monitoring environments. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a mind map for the present invention.

[0016] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1, please refer to Figure 1 As shown in this embodiment, a smart traffic monitoring method based on the Internet of Things includes: Collect multi-source traffic operation information, including vehicle trajectory data, navigation route data, traffic flow data, and road environment data; Based on the comparative analysis of vehicle trajectory data and navigation path data, combined with real-time traffic conditions and passenger behavior characteristics, behavioral intent parameters used to characterize the rationality of vehicle driving behavior are calculated. The original trajectory features and behavioral intent parameters are jointly embedded into the input vector of the deep trajectory clustering model to construct the trajectory behavior vector space; A deep trajectory clustering model is used to perform clustering analysis on trajectory behavior vectors, and the clustering boundaries are dynamically adjusted based on road type, time period and vehicle attribute context elements to construct elastic clustering regions for identifying potential abnormal trajectories. The warning results are compared with the subsequent actual law enforcement results, and the weights of the behavioral intent parameters and the structure of the deep trajectory clustering model are updated online.

[0019] Specifically, this involves collecting multi-source traffic operation information through IoT sensing devices deployed in the road traffic environment, including: The IoT sensing devices include, but are not limited to, the following types of devices, which can be divided into fixed deployment devices and mobile deployment devices: Fixed deployment equipment specifically includes: High-definition video cameras: installed at key locations such as road intersections, main roads, and entrance / exit ramps, to collect traffic flow images, identify license plates, lane occupancy, and vehicle speed; Geomagnetic vehicle detector: Embedded in the road surface, used to sense the frequency of vehicle passage, time intervals, and duration of dwell time; Radar and laser ranging equipment: used to measure vehicle speed, distance between vehicles, and changes in flow direction; Micro-meteorological monitoring device: collects environmental factors such as temperature, humidity, visibility, and wind speed along the road section; Road-sensing landmarks or RFID identification points: used for local vehicle identification and trajectory reconstruction.

[0020] Mobile deployment devices specifically include: On-Board Unit (OBU): Installed on operating vehicles, ride-hailing vehicles, buses or logistics vehicles to collect vehicle GPS trajectory, navigation route, driving status and in-vehicle environment data; Smartphone terminal: By connecting with the driver's or passenger's APP, it uploads information such as real-time navigation routes, user operation behavior, and ride request trajectory; Drones or unmanned patrol vehicles: used for high-angle, wide-area data collection of traffic conditions and emergencies.

[0021] The multi-source traffic operation information collected by IoT sensing devices includes, but is not limited to, the following data types: Vehicle trajectory data: real-time GPS location information (latitude and longitude, timestamp); driving speed, acceleration, and steering angle; historical trajectory sequence of the vehicle (set of location points within a time window); driving behavior characteristics (sharp turns, sudden braking, prolonged idling, etc.); Navigation route data: real-time navigation recommended route; whether the driver follows the navigation (deviation between trajectory and navigation route); navigation change behavior (number of times the route is replanned, reasons for deviation); destination information and starting point information.

[0022] Traffic flow data: vehicle volume per unit time; average vehicle speed and road capacity; congestion index and congestion level; vehicle classification statistics (cars, trucks, buses, motorcycles, etc.); traffic light status and traffic cycle.

[0023] Road environment data: road surface slipperiness, water accumulation monitoring, pothole detection; weather conditions: rainfall, fog and haze, wind speed, temperature, visibility; temporary traffic control information (construction closures, accident-affected areas); static road information such as speed limits, traffic restrictions, and lane divisions.

[0024] The collected multi-source heterogeneous data are uniformly aligned using timestamps to construct a traffic event context based on time series; a map matching algorithm is used to align GPS trajectories with a high-precision road network to reconstruct the actual driving path; data from fixed and mobile devices are fused according to spatial dimensions, and a spatial data model is established through a distributed GIS coordinate system.

[0025] In the IoT-based intelligent traffic monitoring method provided by this invention, the construction and evaluation of behavioral intent parameters are the core steps in identifying the rationality of atypical trajectories. Specifically, based on the collected vehicle trajectory data and navigation path data, the system dynamically generates multiple custom parameters to characterize the rationality of driving behavior by fusing and analyzing real-time traffic environment information and passenger behavior characteristic data. This section will elaborate on the computational mechanism, logical structure, and semantic meaning of this step.

[0026] First, the system acquires vehicle trajectory data and navigation path data in real time based on an IoT platform. The trajectory data includes a series of spatiotemporal points along the vehicle's actual driving path, with each data point including a timestamp, latitude and longitude, speed, and direction angle. The navigation path data is provided in real time by the navigation system or in-vehicle navigation app, including recommended routes, route update records, and estimated arrival times.

[0027] By comparing and analyzing trajectory data with navigation path data in a spatial manner, the system can identify whether the vehicle deviates during actual driving, the degree of deviation, and the duration of the deviation. This comparison process is implemented using a dynamic map matching algorithm, which combines the road network topology to calculate the shortest feasible path distance and angle deviation from the trajectory point to the recommended path.

[0028] To determine the rationality of trajectory deviation, the system further introduces several calculable and quantifiable behavioral intent parameters. The definitions, reasoning mechanisms, and numerical calculation logic of the five core behavioral intent parameters proposed in this invention will be explained below: Navigation deviation rate measures the difference between the vehicle's actual path and the navigation-recommended path during driving. Key factors considered in this parameter include deviation distance, deviation duration, and deviation frequency.

[0029] Within a unit time window, the system compares all trajectory points with the navigation path. If the spatial distance between a trajectory point and the navigation path exceeds a set threshold, it is considered a deviation. The navigation deviation rate is the proportion of deviation trajectory points to the total number of trajectory points per unit time.

[0030] For example, if 30% of the points in a 10-minute driving trajectory deviate from the navigation route by more than 20 meters, the navigation deviation rate is 0.3. The higher the value, the more the behavior deviates from the normal navigation guidance.

[0031] The congestion avoidance factor is used to assess whether a vehicle's deviation from the navigation path is reasonable in terms of traffic congestion avoidance. The system calls on real-time traffic status data from the road condition platform to analyze whether there are any abnormal events such as congestion, slow traffic, or traffic control on the original navigation path when the vehicle deviates from the navigation path.

[0032] The calculation logic for this parameter is as follows: if the original route experiences medium to high levels of congestion during the time when vehicles deviate, and the deviated route is more efficient in terms of traffic flow (e.g., higher average speed), then a higher avoidance factor score is assigned; otherwise, the score is reduced.

[0033] This factor reflects the vehicle's intention to actively avoid adverse road conditions. Reasonable avoidance behavior should be considered normal rather than abnormal trajectories in the system's judgment.

[0034] The passenger intervention index is used to indicate whether passengers actively participate in route selection when there are unexpected changes in the vehicle trajectory.

[0035] This index can be inferred in the following two ways: (1) In ride-hailing or taxi platforms, if there are behaviors such as passengers changing their destination, making special driving requests, or enabling custom routes in the order interaction records, the system will mark their intervention behavior and include it in the score.

[0036] (2) For ride-sharing platforms, the index score will also be improved when there are user behavior logs of actively marking preferred routes or refusing automatic route recommendations.

[0037] The system assigns a score between 0 and 1 to the passenger intervention index based on the frequency of behavior and the intensity of intervention. The higher the score, the stronger the influence of the passenger's subjective will on the deviation from the trajectory, and therefore the more reasonable it is.

[0038] The temporal rarity score is used to measure the probability of the current trajectory occurring within a corresponding time period, and to determine whether the path is an unconventional route.

[0039] The system performs time-period classification and modeling on historical large-sample trajectory data, statistically analyzes common trajectory patterns at the hourly level, and generates a frequency distribution for each type of trajectory. If the current trajectory has a low overlap with historical high-frequency trajectories within its time period, the score is improved, indicating that the trajectory path is relatively rare.

[0040] This parameter is particularly useful for distinguishing between regular commuting routes and abnormal detour routes during morning and evening rush hours, providing a time-series reference for judging atypical behavior.

[0041] The behavioral consistency score reflects the degree of similarity between the current trajectory and the vehicle's past behavioral patterns. The system constructs a "behavioral feature template" for the vehicle by analyzing its driving trajectory samples over a recent period.

[0042] If the current trajectory is highly consistent with the historical trajectory in terms of route selection, speed changes, and travel time, the score will increase, indicating that the deviation may be driven by personal habits; conversely, if the changes are drastic, the consistency score will decrease.

[0043] This parameter helps to personalize the analysis of vehicle behavior differences and avoids the "one-size-fits-all" logic for judging anomalies.

[0044] The five types of behavioral intent parameters mentioned above constitute the behavioral intent feature vector of the vehicle's current trajectory. The system can use a weighted linear model, decision tree model, or neural network model to comprehensively evaluate each parameter and calculate the trajectory's rationality score, depending on task requirements.

[0045] The weights of each parameter can be automatically updated based on the feedback from false positives and false negatives in the training samples, supporting online learning and adaptive adjustment. The final intent score serves as a pre-judgment basis before trajectory clustering, used to adjust the clustering model's sensitivity to anomalies in that trajectory sample.

[0046] In the IoT-based intelligent traffic monitoring method proposed in this invention, to achieve accurate identification of abnormal vehicle trajectory behavior, after collecting vehicle trajectory data and behavioral intent parameters, the two need to be effectively integrated and input into a deep trajectory clustering model for behavior analysis and classification. To this end, this invention proposes a technical approach to construct a trajectory behavior vector space, namely, by jointly embedding the original trajectory features and behavioral intent parameters into a unified model input vector, high-dimensional modeling and deep clustering analysis of trajectory behavior are achieved.

[0047] First, the collected vehicle trajectory data undergoes time-series modeling. This trajectory data contains a sequence of latitude and longitude points traversed by the vehicle within a certain time interval, with each trajectory point having a distinct timestamp. Based on this, the system extracts multiple behavioral features of the vehicle in the spatiotemporal dimensions, including but not limited to the following: Trajectory point sequence: Continuous location data of a vehicle within a unit of time period, reflecting its travel path; Speed ​​change information: including speed values ​​between points and speed difference sequences, used to determine driving stability; Acceleration mode: Calculate the first derivative of the velocity and perform window sliding processing to extract rapid acceleration or deceleration behavior; Turning frequency: Identify vehicle steering behavior by analyzing the angle changes of continuous trajectory points and count the number of turns; Parking behavior recognition: Based on the length of time the speed approaches zero and the distribution of locations, identify whether the vehicle has behavior patterns such as short-term stops or long-term parking.

[0048] The aforementioned features constitute a set of spatiotemporal behavioral features of the vehicle during this period. Since this set exhibits typical time-series characteristics, this invention employs recurrent neural networks (such as LSTM or GRU) or autoencoder models to vectorize this type of time-series data.

[0049] The autoencoder extracts the core behavioral features of the input sequence by compressing and encoding it, and finally generates a set of fixed-dimensional trajectory base encoding vectors. These vectors contain the trajectory behavior patterns and dynamic driving features of the vehicle in that time period, and have high recognizability and input capability.

[0050] To further enhance the ability of trajectory clustering models to understand the semantics of driving behavior, this invention introduces multiple behavioral intent parameters obtained from the aforementioned calculations on the basis of generating basic trajectory encoding vectors, and fuses and spatiotemporally trajectory features to form a multi-dimensional joint feature representation.

[0051] The behavioral intent parameters include, but are not limited to: Navigation deviation rate: Indicates the degree of spatial deviation between the vehicle and the recommended navigation route; Congestion avoidance factor: assesses whether trajectory deviation is caused by reasonable congestion avoidance behavior; Passenger Intervention Index: Reflects whether passengers actively request to change their routes; Temporal rarity score: indicates the rarity of the current trajectory path within that time period; Behavioral consistency score: Determines the degree of consistency between the current trajectory and historical trajectories.

[0052] To achieve the fusion of different data types, this invention normalizes the aforementioned behavioral intent parameters, unifying their value range to the [0, 1] interval. Then, this set of parameters is used as a semantic feature vector and concatenated with the trajectory base encoding vector along the feature dimension to construct the final joint feature representation vector.

[0053] This joint feature representation includes both dynamic behavioral information of the trajectory and "behavioral semantics" behind the current path, such as congestion avoidance intentions and task specificity, and is a multi-level expression of vehicle path behavior.

[0054] After constructing the joint feature representation, the system inputs it into a preset deep trajectory clustering model. The model can be a deep embedding clustering model, a trajectory clustering model based on graph neural networks (such as TrajGAT), or a hybrid structure model using a self-supervised clustering mechanism.

[0055] The main task of clustering models is to map the high-dimensional joint features of the input to the trajectory behavior vector space, such that in this space: Samples with similar trajectory behavior characteristics and similar behavioral intention parameters are relatively close in space, forming high-density clusters; Trajectory samples with different path semantics or atypical behavioral intentions exhibit greater spatial distance and are easier to identify.

[0056] This vector space has both semantic constraints on trajectory behavior and driving intention, and is therefore called the trajectory behavior vector space.

[0057] This invention specifically sets the clustering model to have dynamic learning capabilities. Through subsequent feedback mechanisms (such as manual confirmation and law enforcement results), the embedding weights and cluster boundaries are iteratively optimized to ensure that the space can more accurately distinguish between "reasonable deviations" and "abnormal trajectories" during continuous operation.

[0058] To enhance the flexibility and intelligence of trajectory clustering models in intelligent transportation monitoring systems, this invention, based on a deep trajectory clustering model for identifying vehicle trajectory behavior, introduces a dynamic clustering boundary adjustment mechanism based on context-aware weights. This constructs an elastic clustering region that can adapt to changing scenarios. This region allows for dynamic adjustment of the deviation tolerance of trajectory behavior under different contextual factors, significantly improving the system's accuracy in identifying atypical but reasonable behaviors and reducing the risk of false alarms.

[0059] First, the system constructs a deep trajectory clustering model to analyze and classify the trajectory behavior features of vehicles. This model can be a deep embedded clustering model (DEC), a graph neural network clustering model (such as TrajGAT), or an unsupervised clustering network using a variational autoencoder structure.

[0060] The input to this model is the trajectory behavior vector that has been constructed as described above, which is a multi-dimensional feature vector generated by fusing the following two types of information: Vehicle spatiotemporal trajectory coding features: generated by encoding vehicle trajectory data through recurrent neural networks (RNN, LSTM) or autoencoders, representing its driving path, speed changes, steering patterns, parking behavior, etc. Normalized behavioral intent parameters: including semantic parameters such as navigation deviation rate, congestion avoidance factor, passenger intervention index, time rarity score, and behavioral consistency score, which have been normalized to a uniform scale.

[0061] The deep clustering model embeds and maps the behavior vector, projecting it into a high-dimensional feature space. Trajectories with similar features and behavioral semantics in the space are grouped into the same cluster. At this stage, the model forms multiple basic trajectory clusters and calculates the basic boundary structure for each cluster, such as its center vector, boundary range, and intra-cluster sample density.

[0062] To achieve dynamic adjustability of trajectory clustering boundaries, this invention introduces a context-aware mechanism, which identifies key environmental factors affecting the rationality of trajectories, mainly including: Road type information: such as urban arterial roads, rural roads, highways, residential roads, etc.; Time period information: such as weekday morning rush hour, evening, weekends, holidays, etc.; Vehicle attribute information: such as taxis, ride-hailing vehicles, private cars, police vehicles, cold chain logistics vehicles, etc.

[0063] The aforementioned contextual elements can be obtained in real time through road tags, traffic platform API interfaces, vehicle registration databases, and other means within the IoT traffic sensing network. The system combines and models various types of contextual information to construct a context-aware weight matrix, which indicates the tolerance level for trajectory feature judgment under different scenarios.

[0064] For example, in residential areas at night, the probability of vehicle deviations being considered reasonable is low; while during the morning rush hour, vehicles are more likely to detour to avoid congestion, and deviations are more tolerable. Therefore, the system needs to dynamically adjust the trajectory determination criteria according to changes in context.

[0065] After constructing the basic trajectory clusters, the system extracts boundary features for each cluster, including: Center vector: Represents the typical behavioral pattern of this cluster; Boundary radius: The maximum or average Euclidean distance from a sample in the cluster to the center vector; Sample density distribution: The density of samples within a cluster is calculated using kernel density estimation or the local outlier factor (LOF). Dimensional deviation threshold: For each feature dimension, the standard deviation or maximum offset range of samples within the cluster is calculated.

[0066] The system then maps the context-aware weight matrix to the aforementioned boundary features one by one. For example, in the context combination of "taxi + night + urban side road", a lower tolerance coefficient is assigned to the "navigation deviation rate" dimension, while a higher tolerance weight is assigned to the "passenger intervention index" dimension.

[0067] Based on the mapping results, the system calculates the resilience coefficient of each cluster on each feature dimension. This resilience coefficient reflects the extent to which each feature deviates from the allowed range of expansion in the current context. The resilience coefficient is a real number greater than or equal to 1; a larger value indicates a higher tolerance.

[0068] The system uses the aforementioned elastic coefficients to expand the original cluster boundaries in a non-uniform direction, generating elastic cluster boundary regions with anisotropic variable radii and density thresholds.

[0069] In the distance space, the boundary radius will expand according to the feature dimensions, and the expansion magnitude of each dimension is determined by its corresponding elasticity coefficient. In the density space, if the context indicates that the deviation is acceptable in the case of low density, the anomaly detection density threshold of the cluster should be appropriately lowered. The clustered regions formed in the embedded space exhibit asymmetric structures such as ellipsoids and irregular curved surfaces, which truly reflect the ambiguity of the boundaries of reasonable behavior in real-world scenarios.

[0070] The resulting elastic clustering regions possess the following properties: For typical trajectory samples, the model still uses high-precision clustering; For trajectories located in the boundary area, if their behavioral intention score is within the normal range and does not exceed the extension zone, they are judged as "reasonable deviation" trajectories. The system only identifies a trajectory as a "potentially abnormal trajectory" and triggers an anomaly warning mechanism when the trajectory behavior vector falls outside the elastic boundary or when its multidimensional behavioral intent parameters simultaneously show a high risk of anomaly.

[0071] To further improve the discrimination accuracy and model stability of intelligent traffic monitoring systems in complex trajectory behavior recognition tasks, this invention proposes a closed-loop adaptive optimization mechanism based on actual law enforcement feedback. By continuously collecting deviation information between system warning results and actual law enforcement conclusions, the weight coefficients of behavioral intent parameters are dynamically adjusted, and the structure of the deep trajectory clustering model is fine-tuned online, thereby enhancing the system's learning ability, correction ability, and scenario adaptability.

[0072] After the system completes trajectory behavior vector analysis and elastic clustering region matching, it will immediately or periodically trigger an alert for samples identified as abnormal trajectories. This alert information will serve as the basic recording unit for subsequent judgment and learning optimization.

[0073] To achieve a closed-loop optimization mechanism, the system synchronously or asynchronously collects actual law enforcement outcome data related to the warning event. These data sources include, but are not limited to: Manual review results: Supervisory personnel conduct manual confirmation based on system alerts; On-site enforcement records: Real-time evidence collection and handling results of specific vehicle behaviors in the traffic management system; Behavioral verification information: Post-event rationality analysis based on historical trajectory, task information, and user instructions; User complaint feedback: Objections to abnormal warnings initiated by the driver or passenger.

[0074] The system compares the warning conclusions (such as anomaly markers) of each trajectory sample with the actual law enforcement conclusions (such as whether a violation actually occurred) to form standardized "hit / false alarm marker pairs." That is: If the system alert indicates an anomaly, and law enforcement feedback confirms the existence of abnormal behavior, it will be marked as "hit". If the system alert indicates an anomaly, but law enforcement feedback suggests the behavior was reasonable, it will be marked as a "false alarm".

[0075] The aforementioned labeled pairs, together with the behavioral intent score of the trajectory sample, will form a new training data item for model backtracking and parameter updates.

[0076] The system inputs the continuously accumulating sample-label dataset into the parameter importance analysis module to identify the influence and degree of bias of each behavioral intent parameter in the anomaly judgment process.

[0077] For the hit samples that were correctly identified as anomalies, the average score and standard deviation of each parameter dimension were calculated to analyze their typical score range. If a certain parameter is concentrated in the hit samples, it indicates that it has a strong discriminative ability.

[0078] For samples that are misclassified as anomalous, the system analyzes the degree of deviation of each parameter score relative to the correct samples. If a parameter frequently shows abnormally high (or abnormally low) values ​​in the false alarm samples, it indicates that it has a high sensitivity or distortion effect in that sample set.

[0079] The system can also calculate the marginal impact of parameters on the final judgment result, and combine the changes in cluster boundary distance or the fluctuation of classification score to quantify the degree of perturbation of each dimension parameter on the model output.

[0080] Through the above comprehensive analysis, the system identifies the "highly reliable" parameter dimensions that play a key role in the current sample set, as well as the "highly interfering" parameter dimensions that cause misleading effects in misjudged samples, providing a basis for subsequent weight adjustments.

[0081] Based on the aforementioned importance assessment results, the system adaptively and dynamically adjusts the weight coefficients of the behavioral intent parameters. The specific method is as follows: Weight updates can be based on one or more of the following strategies: Weighted moving average method: Updated based on the moving average between historical weights and the statistical weights of the new sample; Gradient descent strategy: With the hit / false alarm label as the target output, the weights of each parameter are fine-tuned through error backpropagation; Reinforcement learning algorithm: It uses weight adjustment as a policy variable and optimizes the weight configuration policy by using model accuracy as a reward signal.

[0082] If a certain parameter shows an abnormally high value in the false alarm samples and is significantly correlated with the increase in the false alarm rate, the system will reduce the weight of that parameter and weaken its contribution to trajectory anomaly determination. If a certain parameter maintains stability and high discriminative power in the hit samples, the system will increase the weight of that parameter to enhance its semantic expressive power.

[0083] The above update process is executed online, ensuring continuous optimization of weight configuration during actual system operation, so that the overall semantic expression of the behavioral intent vector is closer to the actual behavioral evolution trend.

[0084] Based on the updated behavioral intent parameter weights, the system further fine-tunes the structure of the deep trajectory clustering model to correct the model's embedding representation of trajectory behavior and classification boundaries. This process includes the following: The system remaps all trajectory samples with real feedback labels to the trajectory behavior vector space, which are then used as high-confidence training samples to input into the deep clustering model.

[0085] By incorporating feedback samples, a joint loss function optimization mechanism is introduced to simultaneously minimize the following two error objectives: Cluster boundary error: ensures that samples of the same type are grouped into the same cluster; Intent parameter bias error: Optimize the model's accuracy in identifying high-value behavioral intent dimensions.

[0086] Structural optimization can be achieved in the following ways: The weights of the embedding layer structure are fine-tuned to optimize the projection of trajectory samples into the embedding space; The cluster center vectors are repositioned to better align with the centers of the actual behavior categories; Adjust the boundary expansion parameters of the clusters to optimize the boundary sensitivity for anomaly trajectory recognition.

[0087] If the system introduces a large number of new behavioral pattern samples, a transfer learning strategy can be adopted to incrementally fine-tune the model based on the existing model structure without complete retraining, thereby effectively improving the model's adaptability.

[0088] Example 2, please refer to Figure 2 As shown in this embodiment, a smart traffic monitoring system based on the Internet of Things includes: The data acquisition module collects multi-source traffic operation information, including vehicle trajectory data, navigation path data, traffic flow data, and road environment data. The behavioral intent parameter calculation module calculates behavioral intent parameters to characterize the rationality of vehicle driving behavior based on the comparative analysis of vehicle trajectory data and navigation path data, combined with real-time traffic conditions and passenger behavior characteristics. The trajectory behavior vector space construction module jointly embeds the original trajectory features and behavior intent parameters into the input vector of the deep trajectory clustering model to construct the trajectory behavior vector space; The analysis and adjustment module uses a deep trajectory clustering model to perform clustering analysis on trajectory behavior vectors, and dynamically adjusts the clustering boundaries based on road type, time period and vehicle attribute context elements to construct elastic clustering regions for identifying potential abnormal trajectories. The update module compares the early warning results with the subsequent actual law enforcement results and updates the weights of the behavioral intent parameters and the structure of the deep trajectory clustering model online.

[0089] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A smart traffic monitoring method based on the Internet of Things, characterized in that: include: Collect multi-source traffic operation information, including vehicle trajectory data, navigation route data, traffic flow data, and road environment data; Based on the comparative analysis of vehicle trajectory data and navigation path data, combined with real-time traffic conditions and passenger behavior characteristics, behavioral intent parameters are calculated to characterize the rationality of vehicle driving behavior. The original trajectory features and behavioral intent parameters are jointly embedded into the input vector of the deep trajectory clustering model to construct the trajectory behavior vector space; A deep trajectory clustering model is used to perform clustering analysis on trajectory behavior vectors, and the clustering boundaries are dynamically adjusted based on road type, time period and vehicle attribute context elements to construct elastic clustering regions for identifying potential abnormal trajectories. The warning results are compared with subsequent actual enforcement results, and the weights of behavioral intent parameters and the structure of the deep trajectory clustering model are updated online. The behavioral intent parameters include navigation deviation rate, congestion avoidance factor, passenger intervention index, time rarity score, and behavioral consistency score. A deep trajectory clustering model is used to cluster the trajectory behavior vectors, and the clustering boundaries are dynamically adjusted based on road type, time period, and vehicle attribute contextual elements. Specifically: The deep trajectory clustering model is used to receive trajectory behavior vectors as input, and the trajectory behavior vectors include the spatiotemporal trajectory encoding features of the vehicle and normalized behavioral intent parameters. Through a deep embedding mechanism, multiple input vectors are mapped to a high-dimensional feature space, and the samples in this space are initially clustered to form multiple basic trajectory clusters. Collect contextual element information related to the trajectory, including road type information, time period information, and vehicle attribute information; construct a context-aware weight matrix based on different context combinations, and apply this matrix to the boundary representation of the basic trajectory cluster; Based on the context-aware weight matrix adjustment results, elastic clustering regions are generated. If a trajectory behavior vector is outside the boundary of the main cluster but still within the elastic extension zone, it is not initially classified as an abnormal trajectory. An abnormal trajectory alarm mechanism is triggered only when the trajectory sample exceeds the elastic zone boundary, or when multiple behavioral intent parameters simultaneously exhibit a high risk of anomaly. The elastic clustering regions generated based on the context-aware weight matrix adjustment results specifically include: For each trajectory cluster generated by the deep trajectory clustering model, extract its boundary feature information, including the cluster center vector, boundary radius, intra-cluster sample density distribution, and deviation threshold of feature dimensions; Based on the pre-generated context-aware weight matrix, weight adjustment factors are assigned to the key boundary feature dimensions in different clusters. The weight matrix is ​​dynamically generated according to the road type to which the vehicle belongs, the current time period, and the vehicle attribute category. Based on the mapping relationship between the weight matrix and the boundary features, the elasticity coefficient of each cluster in each dimension is calculated to represent the acceptable deviation range of that dimension in the current context. By using elasticity coefficients to non-uniformly expand the basic cluster boundaries, elastic cluster boundaries with anisotropic variable radii and density thresholds are constructed, forming the final elastic cluster regions.

2. The intelligent traffic monitoring method based on the Internet of Things according to claim 1, characterized in that: The original trajectory features and behavioral intent parameters are jointly embedded into the input vector of the deep trajectory clustering model to construct the trajectory behavior vector space, specifically including: Time series modeling is performed on the collected vehicle trajectory data to extract spatiotemporal behavioral features, including trajectory point sequences, speed changes, acceleration patterns, turning frequencies, and stopping behaviors. Recurrent neural networks or autoencoders are then used to vectorize the spatiotemporal behavioral features to generate trajectory base coding vectors. The calculated behavioral intent parameters are normalized and used as additional semantic dimensions, and then fused and concatenated with the trajectory base encoding vector at the feature layer to form a multi-dimensional joint feature representation. Multi-dimensional joint feature representations are embedded as input into a pre-defined deep trajectory clustering model, and the deep trajectory clustering model is used to construct a trajectory behavior vector space with behavioral intent semantic constraints in a high-dimensional space.

3. The intelligent traffic monitoring method based on the Internet of Things according to claim 1, characterized in that: The warning results are compared with the subsequent actual law enforcement results, and the weights of the behavioral intent parameters and the structure of the deep trajectory clustering model are updated online, specifically including: After identifying abnormal trajectories and generating early warnings based on trajectory behavior vectors and elastic clustering regions, actual law enforcement results data related to the early warning event are collected. Enforcement results data include manual review results, on-site enforcement records, post-event behavior verification information, or user complaint feedback; the warning conclusions and enforcement conclusions of each trajectory sample are compared with the labels to form "hit / false alarm" pair; Based on the scoring results of trajectory samples across multiple behavioral intent parameters and the corresponding hit / false alarm pairs, the weight coefficients of each behavioral intent parameter are adjusted. If a parameter is highly sensitive in false positive samples, its weight is reduced; conversely, if it is significantly discriminative in correct samples, its weight is increased. By combining the accumulated feedback label samples, they are re-embedded into the trajectory behavior vector space, and the embedding layer, cluster center vector, and cluster boundary parameters of the deep trajectory clustering model are adjusted.

4. An Internet of Things (IoT)-based intelligent traffic monitoring system, used to implement the IoT-based intelligent traffic monitoring method according to any one of claims 1-3, characterized in that: include: The data acquisition module collects multi-source traffic operation information, including vehicle trajectory data, navigation path data, traffic flow data, and road environment data. The behavioral intent parameter calculation module calculates behavioral intent parameters to characterize the rationality of vehicle driving behavior based on the comparative analysis of vehicle trajectory data and navigation path data, combined with real-time traffic conditions and passenger behavior characteristics. The trajectory behavior vector space construction module jointly embeds the original trajectory features and behavior intent parameters into the input vector of the deep trajectory clustering model to construct the trajectory behavior vector space; The analysis and adjustment module uses a deep trajectory clustering model to perform clustering analysis on trajectory behavior vectors, and dynamically adjusts the clustering boundaries based on road type, time period and vehicle attribute context elements to construct elastic clustering regions for identifying potential abnormal trajectories. The update module compares the early warning results with the subsequent actual law enforcement results and updates the weights of the behavioral intent parameters and the structure of the deep trajectory clustering model online.

Images