Road traffic abnormal state monitoring system
By using multi-source data processing and hidden Markov models to identify abnormal road traffic conditions, the problem of identifying malicious and distracted driving has been solved, achieving accurate early warning and improved traffic safety.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEBI COLLEGE OF VOCATION & TECH
- Filing Date
- 2026-03-25
- Publication Date
- 2026-05-15
AI Technical Summary
Existing road traffic monitoring systems struggle to accurately identify driving intentions when faced with malicious retaliatory driving and distracted driving, leading to ambiguous information output, hesitation in decision-making by management personnel, and distorted warning content, which may cause traffic accidents.
The system employs modules for multi-source data access and preprocessing, vehicle trajectory tracking and state estimation, vehicle interaction relationship extraction, feature engineering and extraction, intent recognition and confidence calculation, risk quantification and graded early warning decision-making, early warning information dissemination, data storage and historical database, and offline training and model update. By combining Hidden Markov Models and Extended Kalman Filters, it can identify vehicle intents in real time and generate graded early warning commands.
It has achieved accurate identification of malicious and distracted driving, generated precise early warning information, reduced the occurrence of traffic accidents, and improved the intelligence level and safety of traffic management.
Smart Images
Figure FT_1
Abstract
Description
Technical Field
[0001] This invention relates to the field of traffic detection technology, and in particular to a road traffic abnormality monitoring system. Background Technology
[0002] The road traffic anomaly monitoring system is a crucial component of modern intelligent traffic management. Integrating technologies such as the Internet of Things (IoT), big data, and AI, it aims to monitor traffic accidents, vehicle malfunctions, road obstacles, and various abnormal situations around the clock, rapidly disseminating early warning information and shifting traffic management from passive response to proactive prevention. The system constructs a comprehensive sensing network, collecting data on vehicle operation and road conditions through roadside sensing devices and mobile terminals. Edge computing nodes enable rapid local analysis, while cloud-based AI models conduct group behavior analysis and trend prediction to accurately identify abnormal events. Subsequently, warnings are pushed to drivers via navigation voice prompts and map pop-ups. Simultaneously, it provides traffic management personnel with functions such as event review, dispatching, and information dissemination, linking relevant resources for collaborative management and effectively improving the level of intelligent traffic management to ensure safe and smooth road traffic.
[0003] However, even on straight roads with good conditions and sensors functioning ideally, malicious retaliatory driving behavior triggered by socio-psychological factors can occur. Because the physical characteristics of this behavior are highly similar to normal behaviors such as distracted driving, the monitoring system is trapped in a causal inference dilemma, unable to decipher the intent. This leads to a chain reaction of ambiguous information output, hesitation in management decisions, and distorted warning content. Ultimately, the system, which should be preventing accidents, may instead become a catalyst for minor hazards to escalate into actual collisions by inducing panic in following vehicles through generalized warnings.
[0004] Therefore, a road traffic abnormality monitoring system is proposed to solve or alleviate the above problems. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a road traffic abnormality monitoring system.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: A road traffic anomaly monitoring system includes a multi-source data access and preprocessing module, a vehicle trajectory tracking and state estimation module, a vehicle interaction relationship extraction module, a feature engineering and extraction module, an intent recognition and confidence calculation module, a risk quantification and graded early warning decision-making module, an early warning information dissemination module, a data storage and historical database module, and an offline training and model update module. The output of the multi-source data access and preprocessing module is connected to the input of the vehicle trajectory tracking and state estimation module. The output of the vehicle trajectory tracking and state estimation module is connected to the inputs of both the vehicle interaction relationship extraction module and the data storage and historical database module. The output of the vehicle interaction relationship extraction module is also connected to the inputs of both the feature engineering and extraction module and the data storage and historical database module. The feature engineering and extraction module's output is connected to the input of the intent recognition and confidence calculation module and the data storage and history database module, respectively. The intent recognition and confidence calculation module's output is connected to the input of the risk quantification and graded early warning decision module and the data storage and history database module, respectively. The risk quantification and graded early warning decision module's output is connected to the input of the early warning information release module and the data storage and history database module, respectively. The early warning information release module's output is connected to an external display device, an in-vehicle terminal, and a traffic management backend. The data storage and history database module is bidirectionally connected to the offline training and model update module and each real-time processing module, respectively. The output of the offline training and model update module is connected to the data storage and history database module.
[0007] Preferably, the multi-source data access and preprocessing module accesses roadside sensor and floating car data in real time to perform time alignment, coordinate unification and outlier removal on the raw data to generate raw trajectory records with timestamps, vehicle identification, lane number, planar coordinates, instantaneous speed, acceleration, heading angle and turn signal status, and outputs them at fixed intervals.
[0008] Preferably, the vehicle trajectory tracking and state estimation module defines a state vector for each vehicle, including planar coordinates, speed, acceleration, heading angle, and yaw rate, and constructs a state transition function using a constant turning rate and acceleration model. It also uses an extended Kalman filter to recursively estimate the original trajectory to output the smoothed position, speed, acceleration, heading angle, and original turn signal state.
[0009] Preferably, the vehicle interaction relationship extraction module searches for candidate interaction pairs based on lane information at each sampling time to calculate the longitudinal distance, lateral distance, relative longitudinal speed, collision time, and headway between vehicle pairs, so as to filter out valid interaction pairs that meet preset conditions, mark the roles of the front and rear vehicles, and output vehicle pair identifiers and relative motion parameters.
[0010] Preferably, the feature engineering and extraction module extracts historical time windows for each effective interaction pair and extracts the single-vehicle abnormal driving features, two-vehicle game interaction features, and environmental rationality features of the preceding vehicle from the window trajectory. It combines the minimum collision time, minimum headway, and normalized speed difference within the window to form a feature vector, and then outputs the feature vector after standardizing it using normalization parameters read from the historical database.
[0011] Preferably, the intent recognition and confidence calculation module loads the pre-trained Hidden Markov Model parameters from the data storage and history database module, and for each valid interaction pair, it maintains its historical feature sequence and uses a logarithmic forward algorithm to recursively calculate the posterior probability of each intent category at the current time, taking the intent category corresponding to the highest probability as the recognition result and outputting the probability as the confidence level.
[0012] Preferably, the risk quantification and graded early warning decision module calculates a comprehensive risk index based on the minimum collision time and minimum headway within the window, and generates corresponding early warning instructions according to the intent category, confidence level, and comprehensive risk index in accordance with preset grading rules, and outputs the early warning instructions.
[0013] Preferably, the early warning information publishing module receives the early warning instruction and, according to the early warning method and target object in the instruction, pushes the early warning information to the roadside variable message sign, vehicle terminal, navigation application, or traffic management backend and records the publishing result.
[0014] Preferably, the data storage and history module centrally stores the raw data, intermediate results, final results, model parameters, and normalization parameters output by all real-time processing modules, provides the required parameters for each real-time processing module, and provides labeled historical data for the offline training module.
[0015] Preferably, the offline training and model update module periodically extracts new feature samples with real intent annotations from the data storage and history module, re-estimates the initial probability, state transition probability, and Gaussian distribution parameters of the Hidden Markov Model under each state, updates the feature normalization parameters, and writes the updated parameters back to the data storage and history module.
[0016] The present invention has the following beneficial effects: This invention accesses multi-source vehicle trajectory data in real time and estimates motion state, extracts vehicle pairs that may interact and their relative motion parameters, constructs composite features representing abnormal driving and interactive game from historical windows, uses a hidden Markov model to infer the current driving intention category and confidence level, generates graded early warning instructions based on real-time risk indicators and issues them in a targeted manner, and updates model parameters periodically based on event feedback. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a structural block diagram of the present invention.
[0019] The diagram shows: 1. Multi-source data access and preprocessing module; 2. Vehicle trajectory tracking and state estimation module; 3. Vehicle interaction relationship extraction module; 4. Feature engineering and extraction module; 5. Intent recognition and confidence calculation module; 6. Risk quantification and graded early warning decision-making module; 7. Early warning information release module; 8. Data storage and historical database module; 9. Offline training and model update module. Detailed Implementation
[0020] 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, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0021] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0023] In the description of this invention, it should be understood that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0024] Furthermore, the terms "first," "second," and "third" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.
[0025] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0026] A road traffic abnormality monitoring system, such as Figure 1 As shown, the system includes a multi-source data access and preprocessing module 1, a vehicle trajectory tracking and state estimation module 2, a vehicle interaction relationship extraction module 3, a feature engineering and extraction module 4, an intent recognition and confidence calculation module 5, a risk quantification and graded early warning decision-making module 6, an early warning information dissemination module 7, a data storage and historical database module 8, and an offline training and model update module 9. The output of the multi-source data access and preprocessing module 1 is connected to the input of the vehicle trajectory tracking and state estimation module 2. The output of the vehicle trajectory tracking and state estimation module 2 is connected to the inputs of the vehicle interaction relationship extraction module 3 and the data storage and historical database module 8, respectively. The output of the vehicle interaction relationship extraction module 3 is connected to the inputs of the feature engineering and extraction module 4 and the data storage and historical database module 8, respectively. The output of the feature engineering and extraction module 4 is connected to the input of the intent recognition and confidence calculation module 5 and the data storage and history database module 8, respectively. The output of the intent recognition and confidence calculation module 5 is connected to the input of the risk quantification and graded early warning decision module 6 and the data storage and history database module 8, respectively. The output of the risk quantification and graded early warning decision module 6 is connected to the input of the early warning information release module 7 and the data storage and history database module 8, respectively. The output of the early warning information release module 7 is connected to the external display device, the vehicle terminal and the traffic management backend. The data storage and history database module 8 is bidirectionally connected to the offline training and model update module 9 and each real-time processing module, respectively. The output of the offline training and model update module 9 is connected to the data storage and history database module 8.
[0027] Among them, the multi-source data access and preprocessing module 1 accesses roadside sensor and floating car data in real time, performs time alignment, coordinate unification and outlier removal on the raw data to generate raw trajectory records with timestamps, vehicle identification, lane number, planar coordinates, instantaneous speed, acceleration, heading angle and turn signal status, and outputs them at fixed intervals.
[0028] Among them, the vehicle trajectory tracking and state estimation module 2 defines a state vector for each vehicle, including planar coordinates, speed, acceleration, heading angle and yaw rate, and constructs a state transition function using a constant turning rate and acceleration model. It also uses an extended Kalman filter to recursively estimate the original trajectory to output the smoothed position, speed, acceleration, heading angle and original turn signal state.
[0029] Among them, the vehicle interaction relationship extraction module 3 searches for candidate interaction pairs based on lane information at each sampling time to calculate the longitudinal distance, lateral distance, relative longitudinal speed, collision time, and headway between vehicle pairs, so as to filter out effective interaction pairs that meet preset conditions, mark the roles of the front and rear vehicles, and output the vehicle pair identification and relative motion parameters.
[0030] Among them, the feature engineering and extraction module 4 extracts the historical time window for each effective interaction pair and extracts the single-vehicle abnormal driving features, two-vehicle game interaction features and environmental rationality features of the preceding vehicle from the window trajectory. It combines the minimum collision time, minimum headway and speed difference normalized value within the window to form a feature vector and outputs it after standardizing the feature vector using normalization parameters read from the history database.
[0031] The intent recognition and confidence calculation module 5 loads the pre-trained Hidden Markov Model parameters from the data storage and history module 8, and for each valid interaction pair, it maintains its historical feature sequence and uses a logarithmic forward algorithm to recursively calculate the posterior probability of each intent category at the current time, taking the intent category corresponding to the highest probability as the recognition result and using this probability as the confidence output.
[0032] Among them, the risk quantification and graded early warning decision module 6 calculates the comprehensive risk index based on the minimum collision time and minimum headway within the window, and generates corresponding early warning instructions according to the intention category, confidence level and comprehensive risk index according to the preset grading rules, and outputs the early warning instructions.
[0033] Among them, the early warning information release module 7 receives the early warning instruction and, according to the early warning method and target object in the instruction, pushes the early warning information to the roadside variable message sign, vehicle terminal, navigation application or traffic management backend and records the release result.
[0034] Among them, the data storage and history library module 8 centrally stores the raw data, intermediate results, final results, model parameters and normalization parameters output by all real-time processing modules, provides the required parameters for each real-time processing module, and provides labeled historical data for the offline training module.
[0035] Among them, the offline training and model update module 9 periodically extracts new feature samples with real intent annotations from the data storage and history module 8, re-estimates the initial probability, state transition probability and Gaussian distribution parameters of the hidden Markov model under each state, updates the feature normalization parameters, and writes the updated parameters back to the data storage and history module 8.
[0036] The road traffic anomaly monitoring system performs the following steps during operation: Step S1: Acquire real-time vehicle trajectory data, estimate the motion state of each vehicle, and obtain the smoothed vehicle trajectory and kinematic parameters; More specifically, for each vehicle entering the monitoring area, a state vector is defined that includes the vehicle's planar coordinates, speed, acceleration, heading angle, and yaw rate. A constant turning rate and acceleration model is used to construct a state transition function to describe the motion evolution of the vehicle in adjacent time steps; The vehicle trajectory is recursively estimated using an extended Kalman filter. A prediction step and an update step are performed at each sampling time to obtain the smoothed position, velocity, acceleration and heading angle, while retaining the turn signal status in the original observation. The prediction step calculates the predicted state and its covariance matrix at the current time based on the optimal state estimate and state transition function of the previous time step. The update step uses the sensor observations at the current time and their observation noise covariance matrix to calculate the Kalman gain, and then corrects the predicted state to obtain the optimal state estimate and its covariance matrix at the current time. The position, velocity, acceleration, and heading angle from the optimal state estimate are output as smoothed kinematic parameters for use in subsequent steps. Step S2: Based on lane information and vehicle position, extract vehicle pairs that may interact, and calculate the relative motion parameters between each vehicle pair; More specifically, at each sampling time, vehicles are grouped based on lane number. For each vehicle, other vehicles are searched in its own lane and adjacent lanes, and vehicles with a longitudinal distance less than a first preset threshold and a lateral distance less than the lane width are selected as candidate interaction pairs. For each candidate interaction pair, calculate the longitudinal distance, lateral distance, relative longitudinal speed, collision time, and headway between the two vehicles. The collision time is defined as the ratio of the longitudinal distance to the difference between the speed of the rear vehicle and the speed of the front vehicle, and the headway is defined as the ratio of the longitudinal distance to the speed of the rear vehicle. Candidate interaction pairs that meet any of the following conditions are retained as valid interaction pairs: longitudinal distance is greater than zero and less than the second preset threshold; absolute value of lateral distance is less than half of lane width; collision time is less than the third preset threshold or headway is less than the fourth preset threshold. At the same time, the roles of the front and rear vehicles in vehicle alignment are determined based on the sign of the longitudinal distance. When the longitudinal distance is positive, the rear vehicle is the reference vehicle and the front vehicle is the target vehicle. Otherwise, the roles are swapped, and the role information of the front and rear vehicles along with the relative motion parameters are output together. Step S3: For each vehicle pair, extract a fixed-length historical time window, and extract composite features representing abnormal driving and interactive game from the trajectory data within the window to form a feature vector; More specifically, for each valid interaction pair, a fixed-length historical time window is extracted with the current moment as the endpoint, and all smooth trajectory points of the following and preceding vehicles within the window are obtained, including the position, speed, acceleration, heading angle and turn signal status at each moment; Extracting abnormal driving features of the preceding vehicle includes: under conditions of smooth road conditions and turn signals off, counting the number of braking events where the longitudinal deceleration of the preceding vehicle exceeds the first deceleration threshold, dividing by the window duration to obtain the braking frequency; calculating the standard deviation of the preceding vehicle's lateral position within the window as the lateral sway amplitude; using the data from the first half of the window to predict the position of the second half through a constant speed model, calculating the root mean square error between the predicted position and the actual position as the trajectory prediction residual; The interaction features of the two vehicles in the game are extracted, including: detecting the moment when the following vehicle intends to change lanes, i.e., when the turn signal is on and the lateral speed exceeds the lateral speed threshold; for each moment when the following vehicle intends to change lanes, determining whether the preceding vehicle moves laterally in the same direction and decelerates at the same time, counting the total time that the condition is met, and calculating the average value of the lateral overlap of the two vehicles at the moment of the lane change intention, and using the product of the two as the lane change blocking factor; determining whether the following vehicle is in the blind spot to the side and rear of the preceding vehicle, the blind spot condition is that the relative azimuth angle is within a preset angle range and the absolute value of the lateral distance is less than the first lateral threshold, counting the cumulative time that meets the blind spot condition within the window as the blind spot occupancy factor; and calculating the cross-correlation coefficient of the longitudinal acceleration of the two vehicles within the window to characterize the degree of antagonism in acceleration changes. Extract environmental rationality features, use the intelligent driver model to calculate the expected acceleration of the vehicle in front at each moment based on the speed of the vehicle in front, the actual distance and speed difference between the two vehicles, and take the average of the absolute values of the difference between the expected acceleration and the actual acceleration within the window as the behavior-environment deviation degree. The extracted braking frequency, lateral sway amplitude, trajectory prediction residual, lane change blocking factor, blind spot occupancy factor, acceleration cross-correlation coefficient, behavior-environment deviation, as well as the normalized values of minimum collision time, minimum headway, and average speed difference between the two vehicles within the window, are used to construct a feature vector. For each component of the feature vector, normalization is performed using the mean and standard deviation obtained from historical data to obtain a standardized feature vector; Step S4: Model the driving intention of the vehicle pair as the hidden state in the Hidden Markov Model, use the feature vector as the observation sequence, and recursively calculate the posterior probability of each intention category at the current time through the forward algorithm, and output the intention category corresponding to the maximum probability and its confidence. More specifically, a hidden Markov model is pre-trained offline, with its hidden states including three categories: normal driving, distracted driving, and malicious driving. Through a large number of historical feature sequences with real intention labels, the model's initial state probability vector, state transition probability matrix, and mean vector and covariance matrix of the multivariate Gaussian distribution followed by the feature vector in each hidden state are learned by maximum likelihood estimation. During online execution, for each valid interaction pair, a sequence of historical feature vectors from the start time to the current time is maintained. The logarithmic forward algorithm is used to recursively calculate the logarithmic forward probability of each hidden state at the current time. The initial value of the logarithmic forward probability is the sum of the logarithm of the initial state probability and the logarithm of the observation probability at the first time. The recursion of subsequent time steps is based on the logarithmic forward probability of each hidden state at the previous time step, the logarithm of the state transition probability, and the logarithm of the observation probability at the current time step, and is calculated using the logarithmic summation exponential technique. Based on the log-forward probabilities of each hidden state at the current time, calculate the posterior probability of each hidden state, that is, the probability that the hidden state at this time is equal to a certain category. Take the hidden state with the largest posterior probability as the intention category at the current time, and use the largest posterior probability value as the confidence output of intention recognition. Step S5: Generate tiered early warning instructions based on intent category, confidence level, and real-time risk indicators, and send them to the corresponding vehicle or management backend. More specifically, a comprehensive risk index is calculated based on the minimum collision time and minimum headway within the window. This index is a weighted sum of two items: the first item is proportional to the degree of proximity of the minimum collision time to the first collision time threshold, and the second item is proportional to the degree of proximity of the minimum headway to the first headway threshold. When the intent category is malicious driving and the confidence level is not lower than the first confidence level threshold, a level one warning is triggered: a directional warning message containing the license plate number is sent to the provocative vehicle in front via a roadside variable message sign or vehicle-to-infrastructure communication device; a reassuring guidance message containing safe driving advice is sent to the victim vehicle behind via an in-vehicle terminal or navigation application, accompanied by a slight vibration prompt; and event details including trajectory playback, feature indicators and intent probability are pushed to the traffic management backend. When the intent category is distracted driving and the confidence level is not lower than the second confidence level threshold, or the confidence level is between the first confidence level threshold and the second confidence level threshold, a level 2 warning is triggered: only a general prompt about the possible distraction of the vehicle in front is sent to the affected vehicle, and the suspicious behavior is recorded in the background database for manual review; When the real-time collision time corresponding to the comprehensive risk index is lower than the emergency threshold, an emergency audible and visual alarm will be triggered regardless of the intention category, reminding both vehicles to take evasive measures. In other cases, no proactive alerts are triggered; the data is simply stored in the log. Step S6: Collect event result feedback and periodically update the parameters of the Hidden Markov Model using the newly labeled data; More specifically, for each event that triggers an alert, the subsequent results are tracked, including whether a collision occurred, the penalty determination after traffic police intervention, and the review labels of the back-end management personnel. These results are used as the true intent at each moment of the event. Regularly collect newly added feature vector samples with real intent annotations, and re-estimate the parameters of the hidden Markov model by combining them with historical data, including the initial state probability, state transition probability, and the mean and covariance matrix of the multivariate Gaussian distribution in each hidden state. At the same time, update the mean and standard deviation of each component used for feature normalization so that the model can adapt to changes in driving behavior. Historical model versions are retained as a reference, and the performance of the new and old models on the validation set is compared to determine whether to enable the updated model.
[0037] In actual operation, the multi-source data access and preprocessing module 1 collects data from roadside sensors and floating cars in real time, performs time alignment and outlier removal on the original trajectory, and provides clean and time-consistent input for subsequent analysis. This initial step ensures that all subsequent processing is based on reliable data and eliminates misjudgments that may be caused by data noise.
[0038] Next, the vehicle trajectory tracking and state estimation module 2 uses extended Kalman filtering to smooth and predict the motion state of each vehicle, obtaining continuous and accurate position, speed, acceleration and heading angle information, while retaining the original turn signal state. This transforms the originally coarse sensor observations into kinematic parameters that can delicately reflect the dynamic changes of the vehicle, making it possible to capture those fleeting abnormal behaviors.
[0039] Based on this, the vehicle interaction relationship extraction module 3 dynamically filters out vehicle pairs that may cause conflict based on lane number and spatial distance, and calculates the longitudinal distance, lateral distance, relative speed, collision time and headway between each pair of vehicles. This step transforms isolated vehicles into a combination with interactive relationships, expanding the system's focus from single-vehicle behavior to the relative motion between vehicles, thereby creating the necessary context for subsequent identification of common targeted and game-theoretic features in malicious behavior.
[0040] Once a valid interaction pair is obtained, the feature engineering and extraction module 4 begins to play a crucial role. It extracts a fixed-length historical time window and mines composite features that reflect social psychological intentions from the trajectory data within the window. For abnormal driving characteristics of the preceding vehicle, such as the frequency of meaningless braking, the amplitude of lateral sway, and the trajectory prediction residual, it can reveal whether the preceding vehicle is making irrational actions without external causes. Such actions are often random and brief in distracted driving, but may be continuous and targeted in malicious provocation.
[0041] The interactive features of the two-vehicle game, including lane-changing blocking factor, blind spot occupancy factor and acceleration cross-correlation coefficient, directly quantify the degree of confrontation between the following vehicle and the preceding vehicle. When the following vehicle attempts to change lanes, the preceding vehicle deliberately moves synchronously and decelerates. This coordinated response is an important indicator of malicious behavior, while the negative correlation of the acceleration of the two vehicles reflects the game state of acceleration and deceleration.
[0042] The environmental rationality feature compares the deviation between the actual acceleration of the vehicle in front and the expected acceleration predicted based on the traffic environment using an intelligent driver model. This further verifies whether the current behavior is driven by external traffic conditions. If the vehicle in front frequently brakes suddenly even when there is no open space ahead, its behavior is more likely driven by internal emotions. These multidimensional features map driving behaviors that are originally highly similar at the physical layer to a vector space that can reflect socio-psychological intentions, providing rich discriminative basis for subsequent intention inference.
[0043] After obtaining the feature vector, the intent recognition and confidence calculation module 5 introduces a hidden Markov model to perform temporal probability inference. This model divides driving intent into three hidden states: normal, distracted, and malicious. It also learns the state transition probability and the distribution pattern of feature vectors in each state through offline training.
[0044] When running online, the module uses a forward algorithm to recursively calculate the posterior probability of each intent category at the current time, and takes the state corresponding to the highest probability as the recognition result, while outputting the probability value as the confidence level.
[0045] This process makes full use of the temporal continuity of driving behavior. A single accidental braking may be classified as distraction by the model, but multiple consecutive braking with game-theoretic characteristics will be given a higher probability of malice, thereby effectively suppressing misjudgment caused by single-point noise and enabling the system to make more reliable distinctions based on temporal accumulation when facing similar physical features.
[0046] With the output of confidence level, the risk quantification and graded early warning decision module 6 begins to refine the processing by comprehensively considering the intent category, confidence level, and real-time risk indicators. This module calculates the comprehensive risk index based on the minimum collision time and minimum headway within the window, and then generates graded early warning instructions according to preset rules. When the model determines malicious driving with a high confidence level, a first-level early warning is triggered, sending a targeted warning message containing the vehicle's identifier to the provocative vehicle. This direct deterrence can break the behavioral chain of the emotionally out-of-control person. At the same time, it sends reassuring guidance information to the victim vehicle, suggesting that it maintain a safe distance and change lanes when the opportunity arises, so as to avoid entanglement with the provocateur and prevent the following vehicle from taking dangerous actions due to anger or panic. The event details are also pushed to the management backend to assist in manual review.
[0047] When the model determines that the driver is distracted or the confidence level is moderate, a level 2 warning is triggered. Only a general alert is sent to the affected vehicle and the suspicious behavior is recorded in the background. This avoids unnecessary interference while preserving the possibility of subsequent tracing.
[0048] When the real-time collision time is below the emergency threshold, the system will trigger an emergency audible and visual alarm regardless of the intent category, ensuring that evasive measures can be executed with absolute priority at the most dangerous moment.
[0049] Finally, the offline training and model update module 9 periodically collects event results, including accident records, traffic police penalty determinations, and manual review labels. These real results are used as supervision signals to re-estimate the parameters of the hidden Markov model and update the mean and standard deviation used for feature normalization.
[0050] This process allows the model to continuously adapt to the evolution of driving behavior. For example, when the manifestation of malicious driving changes over time, newly added labeled data will guide the model parameters to adjust accordingly, thereby maintaining the long-term effectiveness of its recognition capabilities.
[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A road traffic abnormality monitoring system, characterized in that, The system includes a multi-source data access and preprocessing module (1), a vehicle trajectory tracking and state estimation module (2), a vehicle interaction relationship extraction module (3), a feature engineering and extraction module (4), an intent recognition and confidence calculation module (5), a risk quantification and graded early warning decision-making module (6), an early warning information release module (7), a data storage and historical database module (8), and an offline training and model update module (9). The output of the multi-source data access and preprocessing module (1) is connected to the input of the vehicle trajectory tracking and state estimation module (2). The output of the vehicle trajectory tracking and state estimation module (2) is connected to the inputs of the vehicle interaction relationship extraction module (3) and the data storage and historical database module (8), respectively. The output of the vehicle interaction relationship extraction module (3) is connected to the inputs of the feature engineering and extraction module (4) and the data storage and historical database module (8), respectively. The output of the feature engineering and extraction module (4) is connected to the input of the intent recognition and confidence calculation module (5) and the data storage and history database module (8), respectively. The output of the intent recognition and confidence calculation module (5) is connected to the input of the risk quantification and graded early warning decision module (6) and the data storage and history database module (8), respectively. The output of the risk quantification and graded early warning decision module (6) is connected to the input of the early warning information release module (7) and the data storage and history database module (8), respectively. The output of the early warning information release module (7) is connected to the external display device, the vehicle terminal and the traffic management backend. The data storage and history database module (8) is bidirectionally connected to the offline training and model update module (9) and each real-time processing module, respectively. The output of the offline training and model update module (9) is connected to the data storage and history database module (8).
2. The road traffic abnormality monitoring system according to claim 1, characterized in that, The multi-source data access and preprocessing module (1) accesses roadside sensor and floating car data in real time, performs time alignment, coordinate unification and outlier removal on the raw data to generate raw trajectory records with timestamps, vehicle identification, lane number, planar coordinates, instantaneous speed, acceleration, heading angle and turn signal status, and outputs them at fixed intervals.
3. The road traffic abnormality monitoring system according to claim 1, characterized in that, The vehicle trajectory tracking and state estimation module (2) defines a state vector for each vehicle, including planar coordinates, speed, acceleration, heading angle and yaw rate, and constructs a state transition function using a constant turning rate and acceleration model. It also uses an extended Kalman filter to recursively estimate the original trajectory to output the smoothed position, speed, acceleration, heading angle and original turn signal state.
4. The road traffic abnormality monitoring system according to claim 1, characterized in that, The vehicle interaction relationship extraction module (3) searches for candidate interaction pairs based on lane information at each sampling time to calculate the longitudinal distance, lateral distance, relative longitudinal speed, collision time, and headway between vehicle pairs, so as to filter out effective interaction pairs that meet preset conditions, mark the roles of the front and rear vehicles, and output the vehicle pair identifiers and relative motion parameters.
5. A road traffic abnormality monitoring system according to claim 1, characterized in that, The feature engineering and extraction module (4) extracts the historical time window for each effective interaction pair and extracts the single-vehicle abnormal driving features, two-vehicle game interaction features and environmental rationality features of the preceding vehicle from the window trajectory. It combines the minimum collision time, minimum headway and speed difference normalized value within the window to form a feature vector and outputs it after standardizing the feature vector using normalization parameters read from the history database.
6. The road traffic abnormality monitoring system according to claim 1, characterized in that, The intent recognition and confidence calculation module (5) loads the pre-trained Hidden Markov Model parameters from the data storage and history module (8), and for each effective interaction pair, it maintains its historical feature sequence and uses the logarithmic forward algorithm to recursively calculate the posterior probability of each intent category at the current time, taking the intent category corresponding to the highest probability as the recognition result and outputting the confidence level based on that probability.
7. A road traffic abnormality monitoring system according to claim 1, characterized in that, The risk quantification and graded early warning decision module (6) calculates the comprehensive risk index based on the minimum collision time and minimum headway within the window, and generates corresponding early warning instructions according to the intention category, confidence level and comprehensive risk index according to the preset grading rules, and outputs the early warning instructions.
8. A road traffic abnormality monitoring system according to claim 1, characterized in that, The warning information release module (7) receives the warning instruction and, according to the warning method and target object in the instruction, pushes the warning information to the roadside variable message sign, vehicle terminal, navigation application or traffic management backend and records the release result.
9. A road traffic abnormality monitoring system according to claim 1, characterized in that, The data storage and history module (8) centrally stores the raw data, intermediate results, final results, model parameters and normalization parameters output by all real-time processing modules, provides the required parameters for each real-time processing module, and provides labeled historical data for the offline training module.
10. A road traffic abnormality monitoring system according to claim 1, characterized in that, The offline training and model update module (9) periodically extracts new feature samples with real intent annotations from the data storage and history module (8), re-estimates the initial probability, state transition probability and Gaussian distribution parameters of the hidden Markov model under each state, updates the feature normalization parameters, and writes the updated parameters back to the data storage and history module (8).