A traffic incident detection system based on radar-visual fusion

Abstract

This invention relates to the field of traffic control system technology and discloses a traffic event detection system based on radar-visual fusion. The system includes: constructing a dynamic expected flow field baseline determined by the upstream vehicle speed vector; quantifying the degree of disorder in the local flow field based on the deviation of the actual vehicle speed vector from the baseline; determining traffic events based on the spatiotemporal persistence of this disorder; and, in traffic congestion, automatically switching to monitoring statistical anomalies in the duration of continuous vehicle stagnation to identify secondary events within the congestion. This invention filters out a large amount of false alarm information that is not substantially related to traffic operation by changing the triggering condition for event detection from the identification of a single physical entity to the determination of persistent anomalies in the collective behavior of traffic flow. This reduces the system's false alarm rate and resource consumption, and effectively detects abnormal group driving trajectories caused by invisible risks such as road icing or oil spills.

Description

TECHNICAL FIELD

[0001] The present application relates to a traffic event detection system based on radar and vision fusion, and belongs to the technical field of traffic control systems. BACKGROUND

[0002] With the improvement of sensor performance, the amount of data obtained by the system and the coverage range are correspondingly increased, which also brings new technical problems in application: the system needs to process all detected physical entities in the field of view, including road floating objects, vehicle shadows or temporary stops in emergency parking areas, etc. This not only continuously occupies a large amount of backend computing resources, but also contains a large amount of content that is not substantially related to the traffic flow state in the generated alarm information, thereby causing alarm fatigue of the operating personnel.

[0003] To deal with this situation, the usual technical improvement direction is to optimize the recognition algorithm to filter irrelevant alarms by improving the accuracy of target classification, but this processing method requires model training for an endless long-tail scene, and does not change the basic logic of the detection process, i.e. the trigger of the detection behavior is still based on the identification of a single physical entity.

[0004] Specifically, the existing technology mainly has the following deficiencies: 1. The detection process requires identification of physical entities as a prerequisite, so the computing load of the system is directly related to the number of physical entities in the environment. In complex scenarios, there is an imbalance between computing power consumption and effective information output; 2. The alarm mechanism is directly related to the appearance of physical entities, rather than whether the entity causes persistent disturbance to the traffic flow, which makes it difficult for the system to effectively distinguish between real traffic events and harmless environmental disturbances in terms of working principle. Therefore, how to establish a traffic event detection method that directly places the analysis starting point on the overall running state of the traffic flow, and no longer requires identification of various physical entities as a necessary prerequisite, and only declares the occurrence of an event when a persistent disturbance to the traffic flow is confirmed, has become a technical problem to be solved by the present application. SUMMARY

[0005] The present application provides a traffic event detection system based on radar and vision fusion, which mainly aims to solve the problem of excessive consumption of computing resources and easy generation of a large number of irrelevant alarms caused by the detection method in the prior art which takes identification of physical entities as a prerequisite.

[0006] To achieve the above-mentioned purpose, the present application provides a traffic event detection system based on radar and vision fusion for event detection in a monitoring area divided into multiple grids, which comprises:

[0007] A data acquisition module configured to acquire real-time speed vectors of one or more vehicle targets in the monitoring area;

[0008] The expected flow field construction module, connected with the data acquisition module, is configured to perform distance-weighted average operation on the real-time speed vector of the vehicle target in one or more upstream adjacent grids of each grid in the monitoring area in each control cycle to generate a dynamic expected speed vector of each grid;

[0009] The flow field entropy quantification module, connected with the expected flow field construction module, is configured to obtain the real speed vector of each vehicle target, perform vector subtraction operation on the real speed vector and the dynamic expected speed vector of the grid where the vehicle target is located to obtain a speed deviation vector, and sum the square of the length of one or more speed deviation vectors in each grid to obtain a local flow field entropy value quantifying the disorder degree of the traffic flow in the grid; the dynamic event triggering module connected with the flow field entropy quantification module and the static event detection module connected with the data acquisition module;

[0010] The mode switching control module is configured to deactivate the dynamic event triggering module and activate the static event detection module when the average speed of the vehicle targets in the monitoring area is lower than a congestion threshold value preset based on the road design capacity; wherein the dynamic event triggering module is configured to output a traffic event when the area and duration of a cluster area formed by a plurality of high-entropy grids spatially adjacent meet a preset area threshold value and a preset time threshold value at the same time.

[0011] The static event detection module is configured to monitor the duration of continuous stagnation of each vehicle target, and calculate the stagnation time standardization score of each vehicle target based on the statistical mean and statistical standard deviation of the duration of continuous stagnation of all vehicle targets in the monitoring area, and output an internal secondary traffic event of congestion when the stagnation time standardization scores of one or more vehicle targets spatially adjacent continuously meet a preset standardization score threshold value.

[0012] Preferably, the system further comprises an event nature classification module configured to, after the dynamic event triggering module outputs a traffic event, obtain a vector set composed of the speed deviation vectors of all vehicle targets in the plurality of grids forming the cluster area; calculate the covariance matrix of the vector set and perform eigenvalue decomposition to obtain a first eigenvalue and a second eigenvalue ; wherein when the ratio of the first eigenvalue to the second eigenvalue is greater than an orderliness threshold value preset based on statistical calibration, the risk feature of the traffic event is marked as an orderly disturbance; and when the ratio is close to 1, the risk feature of the traffic event is marked as a chaotic disturbance.

[0013] Preferably, the system further includes: a perception confidence assessment module, configured to calculate the consistency between the true velocity vectors of multiple vehicle targets within each grid, and the consistency between the true velocity vectors and the dynamically expected velocity vectors, to generate a confidence index characterizing the perception quality of the data acquisition module in that grid; and a dynamic threshold adjustment module, configured to, when the confidence index of one or more grids is continuously lower than a preset confidence threshold, adjust the trigger threshold of the local flow field entropy value used by the dynamic event triggering module when making event determination in the one or more grids according to the degree of decrease in the confidence index.

[0014] Preferably, the desired flow field construction module is further configured to logically divide the monitoring area into a two-dimensional grid of uniform size, and when performing distance-weighted averaging, to set a monotonically decreasing weight coefficient for the real-time velocity vector of the vehicle target in the upstream neighboring grid that is farther away from the target grid.

[0015] Preferably, the system further includes an event localization module, which is configured to calculate the geometric centroid of the spatial location of the cluster area or multiple vehicle targets that continuously meet a preset standardized score threshold after the dynamic event triggering module or static event detection module outputs a traffic event, and determine the location of the geometric centroid as the location of the traffic event.

[0016] Preferably, the static event detection module is further configured to set a preset standardized score threshold of 3 for the standardized score of the stagnation duration, and when determining secondary traffic events within congestion, it requires that the number of one or more spatially adjacent vehicle targets and the duration of the state are both greater than their respective preset quantity thresholds and time thresholds.

[0017] Preferably, the system further includes: a historical entropy map accumulation module, configured to periodically perform an accumulation operation at a preset low frequency, the operation including multiplying all local flow field entropy values ​​covering the monitoring area generated by the flow field entropy quantization module in the current control cycle by a preset accumulation weighting factor and superimposing them onto the historical entropy map corresponding to the monitoring area, and multiplying all accumulated values ​​of the historical entropy map by a preset time decay factor less than 1; and a traffic bottleneck identification module, configured to identify areas on the historical entropy map whose accumulated values ​​stably exceed a preset bottleneck threshold, and output them as structural traffic bottleneck locations.

[0018] Preferably, the perception confidence assessment module is further configured to automatically switch the method of calculating the confidence index from the spatial correlation of the speed vector between the vehicle target and its neighboring vehicle targets to the temporal correlation of the current real-time speed vector of each vehicle target and its historical speed vector when the vehicle density in the monitored area is detected to be lower than a sparse threshold preset based on road traffic engineering standards.

[0019] Preferably, the system further includes: a historical confidence map generation module, configured to accumulate historical data with time decay effect on the confidence indicators generated by the perception confidence assessment module at different time points at a preset background frequency, so as to generate a historical confidence map characterizing the long-term perception quality of each location within the monitoring area; and a sensor status self-diagnosis module, configured to identify one or more fixed spatial areas on the historical confidence map where the accumulated confidence value is continuously lower than a preset diagnostic threshold, and output a sensor physical status check and maintenance alarm for the area.

[0020] Preferably, the dynamic event triggering module is also configured to adaptively adjust the preset area threshold and preset time threshold based on the road grade information of the monitored area and the information of different time periods within a day; wherein, for highway sections or nighttime periods, a smaller area threshold and a shorter time threshold are used.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. By calculating dynamic expected velocity vectors for locations within the monitoring area to establish a real-time changing expected flow field baseline, and quantifying the degree of disorder in the local traffic flow field based on the deviation of the vehicle's actual velocity vector from this baseline, the occurrence of traffic events is ultimately determined based on the persistence of this disorder index in both time and space dimensions. This complete workflow transforms the triggering condition for traffic event detection from the identification of individual physical entities to the determination of persistent anomalies in collective traffic flow behavior. The focus of event detection is limited to disturbances that have a substantial and persistent impact on traffic operations, avoiding the consumption of detection resources and false alarms caused by a large number of transient, insignificant physical entities in the environment.

[0023] 2. After a traffic incident is confirmed, the indicator used to quantify the degree of disorder in the flow field is obtained by summing the squares of the speed deviations of multiple vehicles in the local area. The spatial location of the peak of this indicator is determined as the location of the incident. Since the value of this indicator is directly related to the intensity of kinetic energy dissipation or disorder disturbance in the local traffic flow, the centroid of its spatial distribution corresponds to the core location of the disturbance source. This allows the severity assessment of the incident and the precise location of its physical location to be unified in the same calculation process and completed synchronously. The traffic control system does not need to initiate additional analysis procedures to classify and locate the incident after the initial alarm.

[0024] 3. The basis for judging traffic incidents is the degree of deviation of the vehicle's driving state from the expected flow field formed by its surrounding traffic environment, rather than the direct identification of specific obstacles' optical or electromagnetic features by sensors. Therefore, for road anomalies that are difficult for sensors to directly perceive and identify, such as small-scale road icing or oil stains, when passing vehicles experience a small but continuous slowdown or directional correction, this method can still identify potential risks affecting traffic flow by capturing the spatiotemporally concentrated disordered flow field in that area. Simultaneously, the velocity vector of each vehicle is compared with the motion state of surrounding vehicles. The calculated expected velocity vector is compared, which provides an inherent consistency verification mechanism for traffic flow data. When the sensor performance deteriorates due to factors such as severe weather and generates a large number of erroneous vehicle target data that do not conform to the laws of motion, the velocity deviations presented by these data lack the clustering and transmission characteristics of real traffic disturbances in spatial distribution. Instead, they appear as random and isolated noise points. The logic of event determination based on spatiotemporal continuity makes it difficult to meet the event triggering conditions when processing such noise-contaminated data, thereby reducing the misjudgment rate under unstable sensor operating conditions and improving the reliability of all-weather operation. Attached Figure Description

[0025] Fig. 1 This is a logic and data flow diagram of the dual-mode detection function of the system of the present invention;

[0026] Fig. 2 This is a diagram of the two-level distributed physical deployment architecture of the system of this invention;

[0027] Fig. 3 This is a sequence diagram illustrating the adaptive switching between dynamic and static detection modes in this invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0029] The traffic incident detection system based on radar-visual fusion claimed in this invention is configured to detect incidents in a traffic monitoring area divided into multiple logical grids. Functionally, the system includes a data acquisition module, a desired flow field construction module, a flow field entropy quantification module, a dynamic event triggering module, a static event detection module, and a mode switching control module. The data acquisition module acquires real-time status data of vehicles within the monitoring area from the radar-visual fusion sensor. The desired flow field construction module calculates a dynamic desired speed benchmark for each location within the area based on the collective behavior of upstream traffic flow. The flow field entropy quantification module compares the actual motion state of each vehicle with this benchmark to quantify the degree of disorder in the local traffic flow. The dynamic event triggering module and the static event detection module operate under two states: smooth traffic and congested traffic, respectively, determining whether an event requiring traffic control system intervention exists based on different physical criteria. The mode switching control module is responsible for automatically starting and stopping the dynamic and static detection modes based on the average speed of the traffic flow.

[0030] In multi-lane urban expressways, where traffic flow is high and conditions are unpredictable, rapid response to traffic incidents is crucial for ensuring traffic efficiency. Existing incident detection methods require identifying every physical entity within the field of view, such as briefly stopped vehicles or floating objects on the road, often consuming significant computational resources and generating alarms that are not substantially related to the overall traffic flow state. To address this, the system claimed in this invention employs a detection workflow that starts with the traffic flow field state as the analysis starting point. It makes judgments by continuously analyzing the behavior pattern of the traffic flow as a whole to determine whether there are persistent abnormal disturbances. Thus, the system's analysis focuses on disturbance sources that have a substantial impact on traffic efficiency, effectively distinguishing between real traffic incidents and harmless environmental interference in its working principle. The system's operation begins with a data acquisition module, which uses roadside radar-visual fusion sensor units to acquire real-time status information of each tracked vehicle target within the monitoring area at a sampling frequency of no less than 10Hz. This information is organized to include the target ID and two-dimensional spatial coordinates. and velocity vector The data structure forms the data foundation for all subsequent analyses. To establish a benchmark that can reflect the current traffic conditions in real time, the expected flow field construction module first logically divides the monitoring area into a uniform two-dimensional grid, such as a 5m x 5m square grid. Then, for any target grid... This module identifies all neighboring grids within a preset distance range (e.g., 50 meters) in its upstream direction during each control cycle. And based on these upstream neighboring grids Real-time velocity vectors of all vehicle targets The target mesh is generated by distance-weighted averaging. The dynamic desired velocity vector at the current moment One specific implementation of this operation is as follows: ,in For grid Average velocity vector of all vehicles, weighting coefficient Is with the target mesh and upstream grid Distance between Functions that are inversely proportional, for example This procedure makes the upstream traffic flow state, which is closer to the target grid, play a greater role in determining the desired speed. As a result, the system constructs a desired flow field that can dynamically change with the upstream traffic waves, providing an adaptive reference system for subsequent anomaly detection.

[0031] To quantify the local disturbances in traffic flow, the flow field entropy quantization module obtains the true velocity vector of each vehicle target within the field of view. And compare it with the dynamic desired velocity vector of the grid where the vehicle target is located. By subtracting the vectors, we obtain the velocity deviation vector. This vector characterizes the degree to which the behavior of an individual vehicle deviates from its local flow field environment. Given that a real traffic event will cause vehicles in the upstream and nearby areas to collectively exhibit evasive behaviors such as deceleration or lane changes, resulting in a series of directionally disordered and significantly magnitude-disordered velocity deviation vectors, this invention defines a local flow field entropy index to characterize this degree of disorder. Its calculation method is to calculate all within a single grid. The summation is performed by taking the square of the magnitude of the velocity deviation vector of each vehicle, i.e. For example, within a grid, the desired velocity vector is (60, 0) km / h. A vehicle decelerates due to an obstacle ahead, its actual velocity vector becomes (40, 0) km / h, and its velocity deviation vector is (-20, 0) km / h with a magnitude squared of 400. Another vehicle changes lanes to the right, its actual velocity vector becoming (50, 5) km / h, and its velocity deviation vector becoming (-10, 5) km / h with a magnitude squared of 125. The local flow field entropy contributed by these two vehicles to this grid is 525. This index amplifies the drastic deviation through squaring. The impact of traffic behavior can reflect the degree of traffic flow disturbance. It should be noted that an isolated, instantaneous high-entropy point may be caused solely by the non-accidental driving behavior of a single vehicle and does not constitute a traffic event requiring a system response. To ensure the reliability of event judgment, the dynamic event triggering module performs spatiotemporal persistence analysis on the real-time local flow field entropy distribution map covering the entire monitoring area generated by the flow field entropy quantification module. The triggering logic of this module requires that the declaration of a traffic event must simultaneously meet preset conditions in both spatial and temporal dimensions. In the spatial dimension, the system uses connected component analysis... The algorithm identifies clusters of spatially adjacent high-entropy grids, each with an entropy value higher than a preset threshold. For example, based on historical data, it identifies clusters of high-entropy grids at the 95th percentile of the entropy distribution under normal traffic conditions and calculates their total area. The spatial condition is satisfied only when this area exceeds a preset area threshold. This area threshold can be determined based on road grade; for example, on urban expressways, it is set to an area no less than twice the length of a standard vehicle to filter out disturbances caused by single-vehicle behavior. In the temporal dimension, the system requires that the duration of the high-entropy cluster area must exceed a preset time threshold, which is set to cover the duration of most non-accidental driving behaviors, such as 3 seconds. Only when both area and duration conditions are met simultaneously does the dynamic event triggering module output a traffic event alarm and determine the geometric centroid of the high-entropy cluster area as the location of the traffic event. This spatiotemporal clustering-based judgment procedure ensures that the system only responds to real traffic flow disturbances that have a certain spatial influence range and are stable in time.

[0032] To further enhance the system's responsiveness under different traffic environments, the dynamic event triggering module is configured to adaptively adjust the aforementioned area and time thresholds based on the road grade information and different time periods within the day for the monitored area it is deployed in. During system initialization, a configuration file containing multiple sets of threshold parameters corresponding to different road grades (e.g., highways, urban expressways) and different time periods (e.g., daytime, nighttime) is loaded. During system operation, the dynamic event triggering module automatically queries and applies the corresponding threshold combination for the current scenario based on its internal clock and preset road grade identifiers. Specifically, for highway sections or nighttime periods, considering higher vehicle speeds or potentially longer driver reaction times, the system uses smaller area and shorter time thresholds to improve the sensitivity to detect potential risks. Considering that in situations where traffic is severely congested, all vehicle speeds approach zero, based on speed... The flow field entropy increase detection mode, which deviates from the standard, becomes less effective due to the lack of effective kinetic energy change signals. However, secondary events occurring within the congested traffic flow (such as vehicle breakdowns) play a decisive role in dissipating the congestion. To compensate for the detection capability under this specific condition, the mode switching control module continuously monitors the average speed of vehicles in the entire area. When this average speed is lower than the congestion threshold preset based on the road design capacity (e.g., for an urban expressway with a design speed of 80 km / h, this congestion threshold can be set to 5 km / h), and this low-speed state lasts for more than one minute, the system automatically disables the dynamic event triggering module and activates the static event detection module. The static event detection module operates based on the statistical law that vehicles in normal congested flow exhibit alternating creeping and stagnation patterns. This module first monitors and records the duration of stagnation for each vehicle in real time, and then dynamically calculates the statistical mean based on the stagnation duration data of all vehicles in the monitored area. with statistical standard deviation Subsequently, a standardized score for the dwell time of each vehicle was calculated. This score measures the degree to which a single vehicle's dwell time deviates from the group average. When there are one or more spatially adjacent vehicle targets whose dwell time standardized scores are consistently greater than a preset standardized score threshold, such as the threshold of 3 commonly used in statistics to judge extremely low probability events, the static event detection module outputs a secondary traffic event within the congestion and positions the high-scoring vehicle as the core of the event. This dual-modal automatic switching architecture design enables the system to perform event detection in the full range of working conditions, including both smooth and congested traffic.

[0033] Furthermore, to provide the traffic control center with a preliminary judgment on the nature of the event, the system may also include an event nature classification module. After the dynamic event triggering module outputs a traffic event, this classification module obtains the speed deviation vectors of all vehicle targets within multiple grids forming a high-entropy cluster region, and constructs a vector set from these two-dimensional vectors. Subsequently, the module calculates the covariance matrix of this vector set and performs eigenvalue decomposition to obtain the first eigenvalue. With the second eigenvalue These two eigenvalues ​​physically correspond to the degree of dispersion of this vector cloud in the primary and secondary directions, respectively; when the ratio of the first eigenvalue to the second eigenvalue... When the ratio exceeds a pre-defined orderliness threshold (e.g., 10) based on statistical calibration, it indicates that the vehicle's deviation behavior is highly concentrated in a single direction, such as collective deceleration along the road direction. The system then labels this traffic event's risk characteristics as an ordered disturbance. When the ratio is close to 1 (e.g., less than 2), it indicates that the vehicle's deviation behavior is distributed in all directions, exhibiting a chaotic state, such as emergency braking accompanied by evasive maneuvers. The system then labels this as a chaotic disturbance. This function adds a qualitative label with decision support value to the main event alarm with minimal additional computation. Furthermore, to improve the system's perception under adverse weather conditions and other reduced conditions... To ensure operational reliability, the system may also include a perception confidence assessment module and a dynamic threshold adjustment module. The perception confidence assessment module generates a confidence index characterizing the perception quality of the data acquisition module within a grid by calculating the consistency between the true velocity vectors of multiple vehicle targets within each grid, and the consistency between these true velocity vectors and the dynamic expected velocity vectors calculated by the expected flow field construction module. For example, mean cosine similarity can be used as this index. In a region where the perception system is functioning normally, this index value will be close to 1; however, if the sensor outputs a large number of velocity directions due to heavy rain or glare, the value will be lower. When the noise data changes abruptly, the confidence index of that area will decrease. The dynamic threshold adjustment module is configured to, when the confidence index of one or more grids remains below a preset confidence threshold (e.g., 0.7), raise the trigger threshold of the local flow field entropy value used by the dynamic event triggering module to determine events in that area, based on the degree of decrease in the confidence index. The logic of this closed-loop feedback mechanism is that when the system identifies a decline in the quality of sensed data in a certain area, it actively raises the judgment standard for that area, responding only to anomalies with higher signal strength, thereby effectively suppressing false alarms caused by sensed noise. To diagnose chronic and recurring structural congestion points in the traffic network, the system may also include a historical entropy map accumulation module and a traffic bottleneck identification module. The historical entropy map accumulation module is configured to periodically perform a background accumulation operation at a preset low frequency (e.g., every minute). This operation includes multiplying the local flow field entropy value covering the entire area generated by the flow field entropy quantization module in the current control cycle by a preset accumulation weighting factor (e.g., 0.01), and superimposing it onto the historical entropy map corresponding to the monitored area. At the same time, all accumulated values ​​of the historical entropy map are multiplied by a preset time decay factor less than 1 (e.g., 0).(999); The traffic bottleneck identification module analyzes the historical entropy map in the background. High entropy values ​​generated by accidental events decay and smooth out over time, while bottlenecks caused by road design or driving behavior patterns that repeatedly exhibit high entropy accumulate at a rate that consistently exceeds the decay rate, thus forming stable high points on the map. When the system identifies certain areas on the map whose accumulated values ​​consistently exceed a preset bottleneck threshold, it outputs that area as a structural traffic bottleneck location, providing data for traffic engineering optimization.

[0034] Example 1: On a highway with continuous curves at night, a small area of ​​thin ice appears on the road surface due to low temperature. Due to its physical properties, this thin ice cannot be directly detected by conventional optical or radar sensors. At this time, the traffic density is moderate, with vehicles continuously passing at speeds of 80 km / h to 100 km / h. For drivers, this thin ice area is also an invisible risk; they only react with slight lateral deviations, unconscious directional corrections, or deceleration when driving on it. Under these conditions, detection systems based on physical entity recognition are unlikely to respond effectively before an accident occurs. When the system using the technical solution of this invention is deployed on this road section, its… The expected flow field construction module continuously generates a dynamic expected velocity vector for each logical grid within the monitoring area based on the stable velocity vector of vehicles traveling in the unaffected upstream area. This vector represents the driving state that a vehicle should have under normal road conditions in this scenario. When the first batch of vehicles begins to enter the grid area containing thin ice, the driver's slight correction causes the vehicle's actual velocity vector to deviate slightly but continuously from the dynamic expected velocity vector. The flow field entropy quantization module then accumulates the squares of the magnitudes of these velocity deviation vectors according to the calculation procedure disclosed in the aforementioned specific implementation, causing the local flow field entropy value of the several grids covering the thin ice area to begin to rise slightly.

[0035] As subsequent vehicles passed through the area, similar unconscious avoidance behaviors caused by slippery road surfaces were repeatedly reproduced. This caused the local flow field entropy value within the corresponding grid area to continuously accumulate and stabilize at a high level. It should be noted that the dynamic benchmark generated by the expected flow field construction module provides a high signal-to-noise ratio reference system for the flow field entropy quantification module, enabling the effective accumulation and separation of subtle collective speed deviations caused by road conditions from background traffic noise. Furthermore, the dynamic event triggering module detected that the clustered regions formed by these spatially adjacent high-entropy grids exceeded the preset area threshold in area, and the duration of this state was [not specified]. If the preset time threshold is exceeded, the system declares a traffic incident and outputs the geometric centroid of the high-entropy cluster region as the location of the potential risk area. This event is not triggered by direct identification of the physical entity of road icing, but rather by shifting the analysis object from the physical entity to the collective behavior of traffic flow. After the event is triggered, the event nature classification module performs covariance matrix analysis on the velocity deviation vector set within the high-entropy cluster region. Since the vehicle's deviation behavior mainly manifests as deceleration along the driving direction and small-scale lateral correction, its velocity deviation vector cloud exhibits an elongated ellipsoidal shape, corresponding to the first eigenvalue... With the second eigenvalue If the ratio is greater than the orderliness threshold, the system marks the risk characteristics of the event as an ordered disturbance and transmits the preliminary judgment to the traffic control center that the disturbance is caused by a change in traffic conditions rather than a violent collision. Based on the event alarm information and its risk characteristic marking, the traffic control center implemented proactive traffic control measures such as speed limits and road condition warnings on the affected road section before the physical collision occurred, and the traffic flow on the affected road section returned to an orderly state.

[0036] Example 2: To objectively verify the effectiveness of the technical solution of the present invention in event detection under complex traffic scenarios, this example adopts an experimental environment based on multi-agent microscopic traffic simulation. The simulation platform is built based on an intelligent driver model and a symmetrical lane-changing model that minimizes the number of right-of-way changes. It can generate background traffic flow that conforms to the statistical laws of traffic engineering on a virtual highway segment with three lanes and a length of 2km. The sensor data used in the experiment is generated by deploying a virtual radar-visual fusion sensor in the simulation environment. The key performance specifications of the virtual sensor are set as follows: maximum detection distance of 250 meters, speed measurement accuracy of ±0.1km / h, data refresh frequency of 10Hz, and random measurement noise conforming to Gaussian distribution can be injected to simulate the uncertainty in the perception process.

[0037] This experiment aims to quantify the performance differences between the technical solution of this invention and the control technology in terms of detection rate and false alarm rate, especially in scenarios involving low-detectability obstacles and non-physical risk sources. To this end, an experimental group employing the complete technical solution of this invention was set up, named the experimental group of this invention, along with two control groups. Control group 1 uses a detection logic based on entity recognition, i.e., when the sensor detects a stationary target with a radar cross-section greater than 0.5 square meters for more than 3 consecutive seconds, an event alarm is output. Control group 2 is used to verify the specific role of the dynamic desired flow field in the solution of this invention. Based on the experimental group of this invention, it replaces the dynamic desired velocity vector with a fixed road design speed limit (100 km / h) as the calculation... The benchmark for calculating speed deviation was established. In the experiment, four types of event scenarios were injected into the simulated traffic flow. Each scenario was independently repeated 100 times, and the detection results for each group were recorded. The experimental process and data recording are as follows: The four types of event scenarios are: Scenario A: A standard-sized car stops due to a malfunction in the main lane; Scenario B: A tire fragment measuring 0.5m x 0.5m is scattered on the main lane; Scenario C: A 10-meter-long area is set on the road surface, with its road friction coefficient reduced by 15% to simulate a small area of ​​oil slick or thin ice; Scenario D: No real events occur, but measurement noise twice the normal level is injected into the sensor model. Performance data for each experimental group under different scenarios are shown in Table 1.

[0038] Table 1: A comparison of performance data for each test group under different scenarios.

[0039] Analyzing the data in Table 1, all three groups effectively detected standard parked vehicles in Scenario A. In Scenario B, due to the small radar cross-section of tire debris, the detection rate of control group 1 dropped to 43%, while the detection basis of the present invention's sample group was the increase in flow field entropy caused by the avoidance behavior of surrounding vehicles, rather than the physical characteristics of the obstacle itself, and its detection rate remained at 99%. In Scenario C, control group 1 completely failed due to the lack of identifiable physical entities, with a detection rate of 0%, while the present invention's sample group achieved a detection rate of 95% by capturing the small, collective speed deviations of vehicles caused by the decrease in road friction coefficient. It should be noted that the detection rates of control group 2 in Scenario B and Scenario C were lower than those of the present invention's sample group, because its fixed speed reference could not filter out... Speed ​​fluctuations in normal traffic flow cause weak speed deviation signals caused by real risks to be submerged by background noise. This comparative data confirms the role of the dynamic expected flow field in improving the system's signal-to-noise ratio. In the noise stress test of scenario D, the spatiotemporal persistence judgment logic used in the sample group of this invention effectively filters out randomly occurring false high-entropy value points that do not have spatiotemporal aggregation, with a false alarm rate of only 1%, lower than the other two control groups. The experimental data confirms that the technical solution claimed in this invention, by changing the judgment basis for event detection from the identification of individual physical entities to the quantitative analysis of the persistent anomalies of collective traffic flow behavior, has a higher detection rate and a lower false alarm rate compared to the control technology in the detection of low-detectability non-physical traffic events.

[0040] Example 3: This example combines Figs. 1 to 3 This section describes a traffic incident detection system based on radar-visual fusion, such as... Fig. 1As shown, the radar-visual fusion sensor data serves as the system's raw input, containing vehicle ID, location, and velocity vector information. This data is first collected by the data acquisition module, then split into two paths: one path sends the real-time velocity vector information to the desired flow field construction module, while the other path provides the calculated regional average velocity to the mode switching control module. The desired flow field construction module generates a dynamic desired velocity based on the upstream traffic flow and sends it to the flow field entropy quantization module. This module further combines the actual vehicle velocity vectors obtained from the data acquisition module to calculate the deviation between the actual velocity and the desired velocity, thereby obtaining the local flow field entropy value characterizing the degree of traffic flow disturbance. The data is fed into the dynamic event triggering module to determine traffic events under smooth flow conditions in the spatiotemporal dimension. On the other hand, it is also used in the perception confidence assessment module, which dynamically adjusts the event triggering threshold of the flow field entropy quantification module by evaluating the quality of the perception data. At the same time, the mode switching control module switches between the dynamic event triggering module and the static event detection module based on whether the average speed is lower than the preset congestion threshold. The static event detection module is responsible for monitoring the duration of continuous vehicle stagnation to identify secondary events within the congestion. Finally, the traffic events output by the dynamic or static modules contain information on the event type, location, and risk characteristics.

[0041] like Fig. 2 As shown, the architecture is physically divided into two parts: the roadside field and the traffic control center. Data communication between the two is achieved through 5G or fiber optic networks. At the roadside field, the radar-visual fusion sensor is connected to the edge computing node through an internal bus. The node is equipped with real-time processing software, which includes a data acquisition module, a desired flow field construction module, a flow field entropy quantification module, and dynamic and static event detection modules. This enables on-site real-time analysis of raw sensor data and edge decision-making for events. At the traffic control center, an application server and a data storage server are deployed. The central management software running on the application server includes event reception and management services as well as a historical data analysis module for traffic bottleneck identification. The data storage server has established a historical entropy map database and a traffic event log database for long-term data accumulation and archiving. The operation and maintenance terminal provides an interface for managers to interact with the entire system.

[0042] like Fig. 3As shown, the participants in this sequence diagram include a data acquisition module, a mode switching control module, a dynamic event triggering module, a static event detection module, and an abstract system state management entity. Its interaction logic includes two core scenarios: When the real-time average speed reported by the data acquisition module remains below 5 km / h for more than 60 seconds, meeting the conditions for switching from smooth traffic to congested traffic, the mode switching control module sequentially sends a deactivation command to the dynamic event triggering module and an activation command to the static event detection module, causing the system to enter static detection mode. Conversely, when the average speed remains above 5 km / h for more than 60 seconds, meeting the conditions for recovering from congested traffic to smooth traffic, the mode switching control module performs the opposite operation, namely, deactivating the static event detection module and activating the dynamic event triggering module, causing the system to enter dynamic detection mode. If the average vehicle speed does not change significantly, the system maintains its current operating mode.

[0043] Example 4: This example provides an offline calibration procedure for key parameters in the dynamic event triggering module. When the system claimed in this invention is deployed on a new urban expressway segment, a parameter calibration process is performed to match the entropy threshold area threshold and time threshold in the dynamic event triggering module with the traffic characteristics of the segment. This process aims to balance the real event detection rate and the false alarm rate to determine a quantified working point. The initial state of this process is defined as follows: the target is the original spatiotemporal trajectory dataset of vehicle targets continuously collected by the radar-visual fusion sensors deployed on the segment within 72 hours. This dataset covers the traffic flow change cycle of the segment on weekdays and weekends. The enabling environment is a computing server capable of performing data playback and batch processing.

[0044] The first step of the calibration procedure is data preparation and annotation. By combining video recordings of road sections, the collected 72-hour data is manually annotated to generate a baseline truth list of traffic events containing precise timestamps and spatial locations. The event types in the list cover vehicle-stopped vehicles, road debris, and traffic congestion. All unannotated time periods are considered normal traffic conditions. The second step of the procedure is to perform parameter scanning and performance evaluation. First, the annotated dataset is input into the desired flow field construction module and flow field entropy quantification module of this invention for batch processing, generating a local flow field entropy value history database that covers the entire dataset and is synchronized with time. Subsequently, a grid search is performed in a three-dimensional parameter space, which is composed of entropy threshold area thresholds and time thresholds. The search range is determined according to road engineering design specifications, with the area threshold ranged from 10 to 50 square meters and the time threshold ranged from 1 to 5 seconds. For each parameter space... For each parameter combination point, the judgment logic of the dynamic event triggering module is used to perform a traversal calculation on the complete entropy historical database, compare the output event alarm list with the baseline truth list, and calculate the real event detection rate and normal state false alarm rate under the parameter combination. The third step of the procedure is to determine the parameter combination. After completing the traversal of the entire parameter space, a series of (FPR, TPR) data points corresponding to different parameter combinations are obtained. These data points are plotted in the subject working feature space, and the parameter combination corresponding to the data point with the closest Euclidean distance to the coordinate point (0, 1) is selected. After the above process, for the urban expressway section in this embodiment, it is finally determined that when the entropy threshold is set to 450, the area threshold is set to 25 square meters, and the time threshold is set to 3.5 seconds, the system obtains a real event detection rate of 98.5% and a normal state false alarm rate of 1.2%. This parameter combination is then set as the business operation configuration for this section.

[0045] Example 5: This example describes the adaptive operating procedure of the perception confidence assessment module under specific boundary conditions. When the system claimed by this invention is applied to low-traffic roads at night or in the early morning, the traffic flow is sparse and the distance between vehicles increases. At this time, the confidence calculation method based on the spatial correlation of the speed vector between a vehicle target and its neighboring vehicle targets is weakened due to the lack of effective reference. To cope with this situation, the perception confidence assessment module is configured to obtain the sparse threshold preset by the road traffic engineering design standard at startup and continuously monitor the real-time vehicle density in the coverage area during operation. When the vehicle density is detected to be lower than the sparse threshold, that is, less than 5 vehicles per kilometer, the module automatically switches the method of calculating the confidence index from the spatial correlation mode to the time series correlation mode.

[0046] In this time-series correlation mode, for any independent vehicle target within the monitoring area, the system establishes a rolling short-term historical queue containing its continuous velocity vectors over the past 2 seconds. Based on this queue, a Kalman filter algorithm is used to predict a desired velocity vector that conforms to the vehicle's own motion inertia in the current control cycle. Subsequently, the system calculates the difference between the actual velocity vector of the vehicle observed by the sensor in the current cycle and the desired velocity vector predicted based on its own historical trajectory. If the deviation is consistently greater than the reasonable maneuver range preset by the vehicle dynamics model, it indicates that the target's perception data has exhibited a jump that does not conform to the motion law. The system correspondingly lowers the confidence index of the target and raises the threshold of the local flow field entropy value required to trigger a real event in the vicinity of the target through a dynamic threshold adjustment module. This adaptive switching procedure enables the system to effectively verify the quality of perception data for individual targets in sparse traffic scenarios.

[0047] Example 6: When the system claimed in this invention is first deployed on a new traffic segment, a standardized pre-calibration and verification procedure is executed to match its internal model parameters with the unique traffic flow characteristics of that segment. The initial data basis of this procedure is to continuously collect and store 72 hours of raw spatiotemporal trajectory data of all vehicles using radar-visual fusion sensors deployed on that segment, and combine this with synchronized video surveillance recordings. All observable real traffic events in this dataset, including their type, occurrence time, and spatial range, are manually labeled to form a benchmark truth dataset for subsequent parameter optimization. The first step of this procedure is to determine the spatial analysis granularity, i.e., the size of the logical grid. For this purpose, the system uses grids of different sizes within a preset... Within a size range of 3 to 10 meters, data slices containing typical vehicle stopping events from the baseline truth dataset are batch-processed in 0.5-meter increments. The peak intensity and signal-to-noise ratio of the local flow field entropy caused by the event are calculated in space under different grid sizes. Finally, a 5-meter × 5-meter grid size that can achieve a balance between maximizing the peak entropy and minimizing the sensitivity to harmless vehicle maneuvers is selected as the fixed configuration for this road segment. Secondly, in order to calibrate the congestion threshold of the static event detection module, the procedure selects multiple traffic flow data segments that are manually marked as congested but without secondary events from the dataset, calculates the statistical distribution of the average vehicle speed during these time periods, and sets the 90th percentile of this distribution, i.e., 6.5 km / h, as the congestion threshold for this road segment.

[0048] Furthermore, to calibrate the orderliness threshold of the event nature classification module, the procedure extracts speed deviation vector sets from the baseline truth dataset for chaotic disturbances caused by multi-vehicle rear-end collisions and ordered disturbances caused by slow traffic at the main road exit ahead. The eigenvalue ratio of the covariance matrix of these two sample sets is then calculated. The statistical distribution of the two distributions was analyzed, and the ratio of 8.5, which could distinguish the two distributions with the highest accuracy, was selected as the judgment criterion for this module. In addition, the system also includes an online self-diagnostic logic for sensor health status. This logic continuously monitors the output of the perception confidence assessment module. When the confidence index of a large continuous area covered by a sensor is continuously lower than a hardware failure threshold close to zero, and the network heartbeat packet of the sensor is interrupted at the same time, the system will generate a physical status check and maintenance alarm for the specific device, rather than treating it as a general low-confidence area. After completing the above multi-stage calibration and verification procedures, all determined parameters, including logical grid size, congestion threshold, and orderliness threshold, are stored as a dedicated configuration file for the specific road segment. The system then uses this set of locally adapted parameter baselines to carry out online traffic event detection tasks.

[0049] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A traffic incident detection system based on fusion of radar and vision for incident detection in a monitored area divided into a plurality of grids, characterized in that, The system comprises: a data acquisition module configured to acquire real-time speed vectors of one or more vehicle targets in a monitoring area; an expected flow field construction module connected with the data acquisition module and configured to, in each control period, perform distance-weighted average operation on the real-time speed vectors of the vehicle targets in one or more upstream adjacent grids of each grid in the monitoring area to generate a dynamic expected speed vector of each grid; a flow field entropy quantification module connected with the expected flow field construction module and configured to acquire real speed vectors of each vehicle target, perform vector subtraction operation on the real speed vectors and the dynamic expected speed vector of the grid where the vehicle target is located to obtain a speed deviation vector, and sum up the square lengths of one or more speed deviation vectors in each grid to obtain a local flow field entropy value quantifying the disorder degree of the traffic flow in the grid; a dynamic event triggering module connected with the flow field entropy quantification module and a static event detection module connected with the data acquisition module; a mode switching control module configured to deactivate the dynamic event triggering module and activate the static event detection module when the average speed of the vehicle targets in the monitoring area is lower than a congestion threshold preset based on the road design capacity; wherein the dynamic event triggering module is configured to output a traffic event when the area and duration of a cluster area formed by a plurality of high-entropy grids that are spatially adjacent meet a preset area threshold and a preset time threshold simultaneously; the static event detection module is configured to monitor the duration of continuous stop of each vehicle target, calculate a stop duration standardized score of each vehicle target based on the statistical mean and statistical standard deviation of the duration of continuous stop of all vehicle targets in the monitoring area, and output an internal secondary traffic event of congestion when the stop duration standardized scores of one or more vehicle targets that are spatially adjacent continuously meet a preset standardized score threshold.

2. The traffic incident detection system based on the fusion of radar and vision according to claim 1, characterized in that, The system further comprises an event property classification module configured to, after the dynamic event triggering module outputs the traffic event, acquire a vector set composed of speed deviation vectors of all vehicle targets in a plurality of grids forming a cluster area; calculate a covariance matrix of the vector set and perform eigenvalue decomposition thereon to obtain a first eigenvalue and a second eigenvalue ; wherein, when a ratio of the first eigenvalue to the second eigenvalue is greater than an orderliness threshold value preset based on statistics, the risk feature of the traffic event is marked as an orderly disturbance; and when the ratio is close to 1, the risk feature of the traffic event is marked as a chaotic disturbance.

3. The traffic incident detection system based on the fusion of radar and vision according to claim 1, characterized in that, The system further comprises a perception confidence assessment module configured to calculate the consistency between the real speed vectors of a plurality of vehicle targets in each grid and the consistency between the real speed vectors and the dynamic expected speed vector to generate a confidence index representing the perception quality of the data acquisition module in the grid; and a dynamic threshold adjustment module configured to, when the confidence indices of one or more grids continuously fall below a preset confidence threshold, adjust the triggering threshold of the local flow field entropy on which the dynamic event triggering module relies for event judgment in the one or more grids according to the degree of decrease in the confidence indices.

4. The traffic incident detection system based on the fusion of radar and vision according to claim 1, characterized in that, The expected flow field construction module is further configured to logically divide the monitoring area into two-dimensional grids of uniform size, and when performing the distance-weighted average operation, set a monotonically decreasing weight coefficient for the real-time speed vectors of the vehicle targets in the upstream adjacent grids farther away from the target grid.

5. The traffic incident detection system based on the fusion of radar and vision according to claim 1, characterized in that, The system further comprises an event positioning module configured to, after the dynamic event triggering module or the static event detection module outputs a traffic event, calculate the geometric centroid of the spatial position of the cluster area or a plurality of vehicle targets that continuously meet the preset standardized score threshold, and determine the position of the geometric centroid as the position of the traffic event.

6. The traffic incident detection system based on the fusion of radar and vision according to claim 1, characterized in that, The static event detection module is further configured to set a preset normalized score threshold value of the stagnation duration normalized score as 3, and require that the number of spatially adjacent one or more vehicle targets and the duration of the state are both greater than respective preset number threshold values and time threshold values when making a congestion inside secondary traffic event determination.

7. The traffic incident detection system based on the fusion of radar and vision according to claim 1, characterized in that, The system further comprises a historical entropy map accumulation module configured to periodically perform an accumulation operation at a preset low frequency, which includes superimposing, onto a historical entropy map corresponding to the monitoring area, all local flow field entropy values covering the monitoring area generated by the flow field entropy quantization module in a current control period after being multiplied by a preset accumulation weight factor, and multiplying all accumulated values of the historical entropy map by a preset time decay factor less than 1; and a traffic bottleneck identification module configured to identify, on the historical entropy map, a region whose accumulated value stably exceeds a preset bottleneck threshold value as a structural traffic bottleneck position and output the same.

8. The traffic incident detection system based on the fusion of radar and vision according to claim 3, characterized in that, The perception confidence evaluation module is further configured to automatically switch the manner of calculating the confidence index from being based on the speed vector space correlation between a vehicle target and its adjacent vehicle targets to being based on the time series correlation between the current real-time speed vector of each vehicle target and its historical speed vector when detecting that the vehicle density in the monitoring area is lower than a sparse threshold value preset based on a road traffic engineering standard.

9. The traffic incident detection system based on the fusion of visual and radar according to claim 3, characterized in that, The system further comprises a historical confidence map generation module configured to perform historical data accumulation with a time decay effect on the confidence indexes generated by the perception confidence evaluation module at different time points at a preset background frequency to generate a historical confidence map representing the long-term perception quality of each position in the monitoring area; and a sensor state self-diagnosis module configured to identify, on the historical confidence map, one or more fixed spatial regions whose accumulated confidence values are continuously lower than a preset diagnosis threshold value, and output a sensor physical state inspection and operation and maintenance warning for the region.

10. The traffic incident detection system based on the fusion of radar and vision according to claim 1, characterized in that, The dynamic event triggering module is further configured to adaptively adjust the preset area threshold value and the preset time threshold value according to the road grade information of the monitoring area and the time period information within a day; wherein a smaller area threshold value and a shorter time threshold value are adopted for a highway section or a night period.

Images