Intersection vehicle existence detection method based on multi-dimensional microwave feature and intelligent fusion

Abstract

The invention discloses an intersection vehicle existence detection method based on multi-dimensional microwave characteristics and intelligent fusion. The method comprises the following steps: S1, multi-source data acquisition: acquiring multi-source data through a microwave radar sensor array, a traffic signal lamp state interface and a pre-stored high-precision map; s2, data preprocessing: preprocessing the radar echoes to form a target track; s3, feature extraction: extracting multi-dimensional features of micro Doppler, polarization, motion, geometry and the like for each track, and splicing the features into feature vectors; s4, generating a preliminary confidence probability, and inputting the vector into a pre-trained machine learning model to obtain the preliminary confidence probability; s5, intelligent decision making is carried out, regularization correction is carried out in combination with the real-time signal lamp state and the high-precision map, and timing sequence smoothing is carried out on the correction probability of continuous multiple frames of the same target to obtain the final existence probability; and S6, outputting a final result, and comparing the final result with a dynamic judgment threshold to generate and output a final judgment result for judging whether the vehicle exists or not.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation sensing technology, and in particular to a method for detecting the presence of vehicles at intersections based on multi-dimensional microwave features and intelligent fusion. Background Technology

[0002] In urban intelligent transportation systems and autonomous vehicle-road cooperative systems, intersection vehicle presence detection is a crucial foundational technology. Its detection results directly serve core applications such as traffic signal control, intersection safety warnings, and traffic flow statistics.

[0003] Current intersection vehicle detection technologies primarily rely on video sensing and infrared sensing. However, both of these technologies have drawbacks. Video sensing typically captures images of the intersection using cameras and employs computer vision algorithms to identify vehicles. However, this technology is limited by ambient lighting and weather conditions; in low visibility conditions, image quality drops drastically, leading to a significant decrease in detection accuracy. Infrared sensing detects vehicles by sensing the difference in thermal radiation between them and the environment. While unaffected by visible light, its penetration is extremely weak. In heavy rain, dense fog, and other adverse weather conditions, its signal attenuation is severe, significantly reducing the detection range. Furthermore, it struggles to distinguish between a vehicle that is generating heat and the road surface that may also be generating heat.

[0004] To address these shortcomings, the industry began exploring the use of microwave radar technology, which can operate in all weather conditions and is unaffected by sunlight or weather. However, applying radar to the scenario of vehicle detection at intersections has exposed a series of new problems.

[0005] 1. Traditional intersection radars primarily rely on target motion information (Doppler effect) for detection. When a vehicle stops at an intersection waiting for a red light, its Doppler velocity is zero. The radar can easily confuse this with static debris on the roadside or bridge expansion joints, leading to missed detections of "disappearing vehicles." This is fatal to signal control algorithms that rely on the continuous presence of vehicles.

[0006] 2. The intersection environment is complex. Radar echoes come not only from vehicles, but also from guardrails, lampposts, trees, and road surface water during rain or snow. These clutters create strong reflection points, severely interfering with the identification of actual vehicles. Summary of the Invention

[0007] The purpose of this invention is to solve the above-mentioned problems by proposing a method for detecting the presence of vehicles at intersections based on multi-dimensional microwave features and intelligent fusion.

[0008] To achieve the above objectives, the following technical solution was adopted: A method for detecting vehicle presence at intersections based on multi-dimensional microwave features and intelligent fusion includes the following steps: S1: Multi-source data acquisition; The system collects microwave radar echo signals and real-time traffic light status information by deploying microwave radar sensor arrays on traffic lights at intersections and using traffic light status interfaces, and also obtains pre-stored high-precision map information of the intersections.

[0009] S2: Data preprocessing; The microwave radar echo signal acquired in step S1 is preprocessed to obtain a standardized multidimensional data cube.

[0010] The multidimensional data cube is processed to obtain a target object composed of multiple scattering point traces, and a target trajectory is formed based on the target object.

[0011] Specifically, the dot processing includes the following sub-steps: S201: On the multidimensional data cube, a constant false alarm rate (CFAR) detection algorithm is used to identify significant scattering points, wherein the trace of each scattering point includes at least its distance, radial velocity, azimuth angle, and intensity information.

[0012] S202: Using the DBSCAN aggregation algorithm, aggregate the traces of the scattering points that are spatially adjacent and have the same motion state to form a target object representing a single physical object.

[0013] S203: Based on the target object in the current frame, use a multi-target tracking algorithm to associate it with the trajectory of historical frames and update its status to form a target track.

[0014] Specifically, the current estimated position and velocity of the target trajectory are output in real time by the state estimator in the multi-target tracking algorithm.

[0015] During the formation of the target track, microwave radar echo signal data segments corresponding to each target track are extracted, and the data segments are associated with the corresponding target tracks.

[0016] The data segment includes at least the slow time series signal, the horizontal and vertical dual-polarization channel signal, and the broadband echo signal corresponding to the three-dimensional data unit where the target track is located.

[0017] The target trajectory has a unique ID and contains at least real-time position coordinates and radial velocity estimates.

[0018] S3: Feature extraction; For each target track formed in step S2, multidimensional feature extraction is performed, wherein the multidimensional features include at least microDoppler features, polarization features, and motion and geometric features.

[0019] Specifically, the process of multidimensional feature extraction is as follows: The extraction of the micro-Doppler features specifically involves: acquiring a slow time-series signal associated with the target trajectory, performing a short-time Fourier transform on it to obtain a micro-Doppler time spectrum, and extracting at least one time-frequency statistical feature value from the micro-Doppler time spectrum as a micro-Doppler feature.

[0020] The extraction of polarization features specifically involves: acquiring the horizontal polarization channel signals associated with the target track. Vertical polarization channel signal The amplitude ratio R and phase difference φ of the two channel signals are calculated respectively, and the polarization characteristics are constituted by the amplitude ratio R and phase difference φ.

[0021] The extraction of motion and geometric features specifically involves: From the tracking status of the target trajectory, obtain the radial velocity v of the current target trajectory and the radar cross section calculated based on the microwave radar echo signal power corresponding to the target trajectory.

[0022] A broadband echo signal corresponding to the target trajectory is acquired, and a one-dimensional high-resolution range image is obtained by pulse compression processing of the broadband echo signal. From the one-dimensional high-resolution range image, the number of peak points exceeding a preset amplitude threshold is extracted as the number of scattering points feature, and the radial distance span from the first peak point to the last peak point in the range image is extracted as the length feature.

[0023] The motion and geometric features include at least one or more of the following: the radial velocity v of the target trajectory, the radar cross section, the number of scattering points, and the length feature.

[0024] The multidimensional features also include energy domain features. The specific extraction process is as follows: extract the one-dimensional range image of the zero Doppler channel of the continuous multi-frame series associated with the target track, construct a time series based on the peak amplitude of the one-dimensional range image, and calculate the variance of the time series. The variance of the time series is inversely proportional to the stability of the echo energy.

[0025] The extracted multidimensional features are normalized and concatenated into a multidimensional feature vector F.

[0026] S4: Generate preliminary confidence probabilities; The multidimensional feature vector F obtained in step S3 is input into a pre-trained machine learning classification model, which outputs a preliminary confidence probability P.

[0027] Specifically, the machine learning classification model outputs a continuous value between [0,1] as the initial confidence probability.

[0028] S5: Intelligent Decision Making; Includes the following sub-steps: S501: Based on the target trajectory, real-time traffic light status information, and pre-stored high-precision intersection map, the preliminary confidence probability P obtained in step S4 is regularized and corrected to obtain the corrected probability p.

[0029] Specifically, the rule-based correction is as follows: based on the real-time traffic light status information and the position and movement status of the target trajectory in the high-precision map of the intersection, the initial confidence probability P is corrected using predefined rules corresponding to the traffic light phase and spatial area to obtain the corrected probability p; The predefined rules include: When the current traffic light phase is red, determine whether the target trajectory is located within a preset waiting area before the stop line of the lane corresponding to the red light phase, and whether the absolute value of the radial speed of the target trajectory is less than a preset first speed threshold; if it is located, apply a first weighting factor greater than 1 to the initial confidence probability P for positive gain correction; if it is not located, apply a second weighting factor not greater than 1 to the initial confidence probability P for correction, or leave the initial confidence probability P unchanged. When the current traffic light phase is green, determine the movement state of the target trajectory within the green light lane; if its radial velocity is greater than a preset second movement threshold, it is considered to be moving, and a third weighting factor of approximately 1 is applied to the initial confidence probability P for correction or the initial confidence probability P is kept unchanged; if its radial velocity is less than a preset third movement threshold, and the duration of this state exceeds a preset time window, it is considered to be stationary or moving at extremely low speed, and a fourth weighting factor less than 1 is applied to the initial confidence probability P for negative suppression correction. When the current traffic light phase is yellow, based on the target track's current position, speed, and remaining traffic light time, predict whether it can cross the stop line before the red light turns on; if it is predicted that it can, then the green light phase passage rules are used for correction; if it is predicted that it cannot, then the red light phase waiting area rules are used for correction. The correction probability p is limited to no more than 1.0.

[0030] S502: Perform temporal smoothing on the corrected probability p of multiple consecutive frames of the same target trajectory to obtain the final existence probability. ; Specifically, the time-series smoothing process involves using a Kalman filter algorithm or an exponentially weighted moving average algorithm to smooth the correction probability p of the same target trajectory in multiple processing cycles that are adjacent in time.

[0031] S503: Preset a decision threshold to determine the final existence probability. The result is compared with the judgment threshold, and based on the comparison result, a final judgment result on whether the vehicle exists is generated; Specifically, the dynamic decision threshold is dynamically set according to at least one of the following strategies: Scene type strategy: Determine the decision threshold based on the predefined detection scene type of the area where the target track is located; Track age: Set an age threshold. If the establishment time of a newly established track is less than the age threshold, the decision threshold is reduced by 10% to 20%. Hysteresis comparison strategy: A first threshold and a second threshold are preset respectively. When the environmental signal-to-noise ratio is lower than the first threshold, the decision threshold is lowered; when the environmental signal-to-noise ratio is higher than the second threshold, the decision threshold is raised; when the environmental signal-to-noise ratio is between the first threshold and the second threshold, the decision threshold of the previous moment remains unchanged.

[0032] S6: Final result output; The final judgment result is compared with the identifier of the target track that produced the result, the lane information, and the final existence probability. Output them together.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extracts and analyzes the micro-Doppler features of a target, enabling the detection of minute vibrations in internal components such as vehicle engines. This provides a reliable basis for distinguishing between "stationary but surviving vehicles" and "inanimate static metal debris," significantly improving the detection accuracy of stationary vehicles before stop lines at intersections. Furthermore, by combining multi-dimensional information such as polarization features, motion, and geometric features, and through the synergy of multiple features, it avoids the limitation of a single micro-Doppler feature being weak or even disappearing due to the loss of engine vibration in completely off vehicles, further increasing the confidence probability of the presence of stationary vehicles.

[0034] 2. This invention utilizes the physical penetration characteristics of microwave signals and further extracts the polarization features of the target to effectively suppress specular reflection clutter interference caused by slippery road surfaces, snow accumulation, and other conditions, thereby ensuring the high reliability and stability of the detection system under complex weather conditions.

[0035] 3. This invention incorporates real-time traffic light status and high-precision map information into the decision-making process, and performs rule-based correction and fusion of the initial perception results based on spatiotemporal context. This makes the system output more consistent with the logic of actual traffic scenarios, reduces misjudgments caused by perception uncertainty, and provides more reliable input for intelligent traffic signal control.

[0036] 4. By designing a dynamic decision threshold adjustment mechanism based on the environmental signal-to-noise ratio, the present invention enables the system to automatically optimize detection sensitivity according to real-time signal quality, maintain a high vehicle detection rate in harsh environments, and improve detection accuracy in good environments, thereby enhancing the overall adaptability and practicality of the system. Attached Figure Description

[0037] Figure 1 This is a flowchart of the steps of the intersection vehicle presence detection method based on multi-dimensional microwave features and intelligent fusion in Embodiment 1 of the present invention. Detailed Implementation

[0038] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0039] The system for implementing this invention includes: Microwave radar sensor array: A frequency-modulated continuous wave radar with a center frequency of 77 GHz and a bandwidth of not less than 1 GHz is selected, equipped with dual-polarized (horizontal H and vertical V) transceiver antennas. The sensor is side-mounted on the traffic light pole at the intersection, and the elevation angle is adjusted downward so that its beam can cover the target lane area. The installation height is usually 5-8 meters.

[0040] Edge computing unit: Deployed within the intersection chassis, equipped with a high-performance processor and AI acceleration module. This unit can establish communication connections with radar sensors and traffic signal controllers via Ethernet interfaces.

[0041] Traffic signal interface: It obtains the light color status and countdown information of each phase from the intersection signal controller in real time through standard protocols.

[0042] Data storage and communication module: This module stores the pre-installed high-precision intersection map and uploads the detection results to the traffic management platform via wired or wireless network. The high-precision intersection map includes information such as lane lines, stop lines, and pedestrian crossing GIS data.

[0043] like Figure 1 As shown, a method for detecting vehicle presence at intersections based on multi-dimensional microwave features and intelligent fusion includes the following steps: S1: Multi-source data acquisition; The microwave radar sensor array deployed on the traffic lights at the intersection continuously collects the original baseband echo signal, i.e., the microwave radar echo signal, at a refresh rate of 100Hz.

[0044] Meanwhile, the traffic signal interface obtains the light color status and countdown information of each phase from the intersection signal controller in real time through standard protocols, i.e., real-time signal light status information.

[0045] S2: Data preprocessing; First, the microwave radar echo signal acquired in step S1 is preprocessed to obtain a standardized range-Doppler-angle three-dimensional data cube.

[0046] Specifically, after receiving the microwave radar echo signal, the edge computing unit executes the following preprocessing chain: Range dimension: After windowing the microwave radar echo signal of each pulse, a Fast Fourier Transform is performed to convert the signal from the time domain to the frequency domain, obtaining range cells. Each range cell contains a complex value whose amplitude represents the sum of the echo intensities of all scatterers within that range cell. The range resolution is determined by the radar bandwidth, which is 0.15 meters in this embodiment.

[0047] Doppler dimension: For the echo signal sequence of multiple consecutive pulses on the same range cell, a Fast Fourier Transform is performed again to generate a range-Doppler (RD) matrix. The rows of the range-Doppler (RD) matrix are range cells, and the columns are Doppler cells. The value of each point in the matrix represents the echo energy at a specific distance and velocity. The velocity resolution is determined by the total observation time, and in this embodiment, it is 0.1 m / s. Angular dimension: Using the receiving antenna array and a digital beamforming algorithm, the azimuth angle of the same data cell in the range-Doppler (RD) matrix is ​​estimated, ultimately forming a three-dimensional data cube containing range, velocity, and angular information. The azimuth angle refers to the horizontal angle between the target's azimuth and the normal direction of the radar antenna array. Each three-dimensional data cell in this data cube represents a potential scattering point and its energy intensity at a specific distance, velocity, and azimuth, providing a fundamental information source for subsequent target detection.

[0048] Next, point processing is performed on each three-dimensional data cell in the three-dimensional data cube to obtain a target object composed of multiple scattering point traces, and a target trajectory is formed based on the target object.

[0049] Specifically, dot processing includes the following sub-steps: S201: On the three-dimensional data cube, an ordered statistical constant false alarm rate (CFPR) detection algorithm is used, setting the false alarm probability to 1e-4, to identify three-dimensional data units with energy significantly higher than the background, i.e., scattering points. Each scattering point's trace includes at least its distance, radial velocity, azimuth angle, and intensity information.

[0050] S202: Subsequently, using the DBSCAN aggregation algorithm, with the neighborhood radius ε=0.5 meters and the velocity tolerance δ=0.5 m / s, the traces of multiple scattering points that are spatially adjacent and have the same motion state are aggregated to form a target object representing a single physical object. Each target object is represented by the centroid position and average velocity of its trace cluster.

[0051] S203: Based on multiple target objects in the current frame, a Kalman filter multi-target tracking algorithm is used for target association and tracking. The state vector of the Kalman filter algorithm is set as [x, y, v_x, v_y], and the observation vector is [R, A_z, V], representing the planar position, velocity, and radar polar coordinate observation values, respectively. The target objects clustered in the current frame are associated and matched with existing target tracks in historical frames. If a match is successful, the track state is updated, thus forming a target track; new targets that do not match are initialized with new tracks. The current estimated position and velocity of the target track are output in real time by the state estimator in the multi-target tracking algorithm. In this embodiment, the Kalman filter algorithm is used for multi-target tracking.

[0052] Furthermore, the target track has a unique ID and includes at least real-time position coordinates and a radial velocity estimate. During the formation of the target track, data segments of the microwave radar echo signal corresponding to the target track are extracted and associated with it. These data segments include at least the slow time-series signal corresponding to the three-dimensional data unit of the target track, horizontal and vertical dual-polarization channel signals, and broadband echo signals. The three-dimensional data unit is determined by the current estimated range and azimuth of the target track.

[0053] S3: Feature extraction; For each target track formed in step S2, multidimensional feature extraction is performed to extract three types of features from the data segments associated with the target track, including micro-Doppler features, polarization features, and motion and geometric features.

[0054] Specifically, the process of multidimensional feature extraction is as follows: Micro-Doppler features: Based on the current distance and angle of the target track, extract the slow time series signal corresponding to the position, and perform a short-time Fourier transform on the slow time series signal to obtain the micro-Doppler time spectrum. Extract time-frequency statistical feature values ​​from the micro-Doppler time spectrum as micro-Doppler features.

[0055] In this embodiment, the time-frequency statistical characteristic values ​​are taken as the spectral entropy and the spectral centroid, wherein the formula for calculating the spectral entropy is: in This represents the normalized distribution of the average power of the spectrum along the frequency axis. A high entropy value indicates a dispersed spectrum, while a low entropy value indicates concentrated energy; the formula for calculating the spectral centroid is... The centroid of the spectrum represents the energy center frequency of the spectrum.

[0056] The polarization characteristics are: acquiring the horizontal polarization channel signals associated with the target track. Vertical polarization channel signal The amplitude ratio R and phase difference φ of the two channel signals are calculated respectively, and the polarization characteristics are constituted by the amplitude ratio R and phase difference φ.

[0057] In this embodiment, the formula for calculating the amplitude ratio R is: The formula for calculating the phase difference φ is: Vehicle targets typically have a specific amplitude ratio R and a phase difference φ close to 0° or 180°.

[0058] The aforementioned motion and geometric features: From the tracking state of the target trajectory (i.e., the state vector maintained by a multi-target tracking algorithm, as described in step S203), obtain the radial velocity v of the current target trajectory and the radar cross section (RCS) calculated based on the power and distance of the microwave radar echo signal corresponding to the target trajectory.

[0059] A broadband echo signal corresponding to the target trajectory is acquired, and a one-dimensional high-resolution range profile is obtained by pulse compression processing of the broadband echo signal. From the one-dimensional high-resolution range profile, the number of peak points exceeding a preset amplitude threshold of -10dB is extracted as the number of scattering points feature, and the radial distance difference between the first peak point and the last peak point in the one-dimensional high-resolution range profile is extracted as the length feature.

[0060] Furthermore, in addition to the features mentioned above, multidimensional features also include energy domain features, the specific extraction process of which is as follows: The range-Doppler (RD) matrix generated in step S2 has its horizontal axis corresponding to different radial velocities. All Doppler elements with an absolute radial velocity less than 0.1 m / s are defined as zero-Doppler channels. These zero-Doppler channels contain echo information from all stationary targets in the scene.

[0061] In this embodiment, ten consecutive frames of zero-Doppler channel one-dimensional range images associated with the target trajectory are extracted. The peak amplitude of each frame's one-dimensional range image is recorded, forming a time series of length 10. The time series variance is then calculated based on this series. A smaller variance value indicates a more stable echo amplitude, suggesting the target is more likely to be a rigid metallic object, such as a stationary vehicle. A larger variance value indicates greater echo fluctuations, suggesting the target is more likely to be a non-rigid object such as vegetation or pedestrians, or clutter. Therefore, the variance of this time series is inversely proportional to the stability of the echo energy.

[0062] The stability metric obtained by inverting the variance of the time series is used as the energy domain feature. The energy domain feature and the vibration-sensitive micro-Doppler feature are essentially complementary, together ensuring high robustness detection of vehicles in all states.

[0063] In this embodiment, the motion and geometric features include the radial velocity v of the target trajectory, the radar cross section (RCS), the number of scattering points, and the length characteristics.

[0064] All the extracted multidimensional features are combined into a vector [H, C, R, φ, v, RCS, number of scattering points, length]. The vector composed of these multidimensional features is normalized using the maximum-minimum normalization method, scaling each feature dimension to the [0,1] interval, and then concatenating them into a single multidimensional feature vector F. The normalization parameters used, i.e., the minimum and maximum values ​​of each feature dimension, are statistically determined and stored from the training set during the machine learning model training phase.

[0065] S4: Generate preliminary confidence probabilities; The multidimensional feature vector F obtained in step S3 is input into a pre-trained machine learning classification model, which outputs a preliminary confidence probability P.

[0066] Specifically, the machine learning classification model outputs a continuous value between [0,1] as the initial confidence probability P.

[0067] Furthermore, the training process of this machine learning classification model is as follows: Radar data covering various scenarios including sunny, rainy, foggy, daytime, and nighttime was collected and simultaneously annotated with video data to construct a dataset containing tens of thousands of "vehicle" and "non-vehicle" samples. Non-vehicle samples typically include pedestrians, bicycles, and debris.

[0068] Perform the above feature extraction process on each sample to generate a set of labeled feature vector samples.

[0069] Training Task and Objective: The training task of this model is a supervised regression task, aiming to learn the mapping relationship from the multi-dimensional feature vector F to the vehicle presence confidence score. The vehicle presence confidence score is manually labeled, with vehicle samples labeled as 1.0, non-vehicle samples labeled as 0.0, and ambiguous samples labeled as intermediate values ​​between 0 and 1. The training objective is to minimize the mean squared error loss between the model's predicted values ​​and the vehicle presence confidence scores.

[0070] Train a gradient boosting decision tree (XGBoost) model using a feature vector sample set. Set the maximum tree depth to 6, the learning rate to 0.1, and train for 100 epochs.

[0071] Convergence Criteria: During training, the dataset is divided into training, validation, and test sets with a weighting of 7:2:1. A convergence criterion is reached when the loss function on the validation set no longer decreases over 20 consecutive training epochs, or when the model's coefficient of determination R on the independent test set reaches a certain level. 2 Once the preset threshold of 0.95 is reached, the model is considered to have converged, training is stopped, and the optimal parameters are saved.

[0072] S5: Intelligent decision-making, including the following sub-steps: S501: Read the real-time traffic light status. Based on the current traffic light phase, the target track's position in the high-precision map, and its radial velocity, correct the initial confidence probability P using predefined rules; The predefined rule is: The preset first speed threshold is 3km / h, the second speed threshold is 5km / h, the third speed threshold is 1km / h, and the time window is 10 seconds.

[0073] If the current phase is red, obtain the list of lanes controlled by that red light phase. Based on the pre-stored high-precision intersection map, determine whether the current position of the target trajectory is within the rectangular waiting area 0-5 meters before the stop line of its lane, and whether its radial velocity absolute value is less than the first speed threshold.

[0074] If the region is located within the area and the radial velocity meets the condition, a weighting factor greater than 1 is applied to the initial confidence probability P for positive gain correction, resulting in a corrected probability p. The weighting factor for this positive gain is set to 1.5, and the specific calculation formula is as follows: The correction probability p is limited to no more than 1.0.

[0075] If the object is located within the region but its speed is not lower than the threshold, or if it is not located within the region at all, then the corrected probability p is equal to the initial confidence probability P.

[0076] If the current phase is green, retrieve the list of lanes controlled by that green phase. Determine if the target track is located within a green lane.

[0077] If a vehicle is located within the lane and its radial velocity absolute value is greater than the second speed threshold, it is considered a vehicle on the journey, and the corrected probability p is equal to the initial confidence probability P.

[0078] If a vehicle is located within the lane but its absolute radial velocity is less than the third velocity threshold and the duration exceeds the time window, it is considered abnormally stationary. A weighting factor less than 1 is applied to P for negative suppression, resulting in a corrected probability p. The weighting factor for this negative suppression is set to 0.6, and the calculation formula is as follows: If the current phase is yellow, calculate the distance from the target track to the stop line, and combine the radial velocity of the target track and the remaining time of the yellow light to predict whether it can pass the stop line before the red light comes on.

[0079] If it passes, it will be processed as a green light; If it cannot pass, it will be treated as a red light.

[0080] In any case, the correction probability p is restricted to the interval [0, 1.0].

[0081] S502: Based on the target track ID, cache the correction probability p of the most recent 10 frames for each target track. Perform temporal smoothing on the correction probability p of the same target track for 10 consecutive frames to obtain the final existence probability. ; Specifically, the temporal smoothing process involves using a Kalman filter algorithm to smooth a sequence consisting of 10 consecutive frames of corrected probabilities p, and outputting the smoothed final existence probability. .

[0082] Furthermore, in this embodiment, the Kalman filter state is set as a probability value, and the principles for setting its process noise covariance and observation noise covariance are as follows: Process noise covariance: Used to assess the drastic change in the correction probability p caused by a change in vehicle presence state. Since vehicle presence is relatively stable over a short period, this change is slow and small; therefore, the process noise covariance should be set to a small value. This can be determined by analyzing the statistical variance of the probability p during vehicle presence state transitions in a large number of real-world scenarios.

[0083] Observation noise covariance: used to construct the uncertainty of the correction probability p. This uncertainty mainly originates from sensor noise and feature extraction fluctuations. The magnitude of the observation noise covariance determines the degree of confidence the filter has in the current correction probability p. The observation noise covariance can be set by calculating the variance of the historical sequence of the correction probability p, based on the volatility of the historical sequence.

[0084] As one feasible embodiment, the initial value of the process noise can be set to 1e-4, and the initial value of the observation noise covariance can be set to 1e-3. In actual deployment, these initial values ​​can be adjusted based on the stability of the noise effect. Those skilled in the art can determine suitable process noise covariance and observation noise covariance without creative effort based on the above principles.

[0085] S503: Dynamically preset a decision threshold, which will determine the final probability of existence. The result is compared with the judgment threshold, and based on the comparison result, a final judgment result on whether the vehicle exists is generated; Specifically, the dynamic decision threshold is dynamically set according to at least one of the following strategies: Scene type strategy: The decision threshold is determined based on the predefined detection scene type of the area where the target track is located. For example, if the area is a main lane, the decision threshold is set to 0.7; if the area is a turning lane, the decision threshold is set to 0.6; if the area is a bus lane, the decision threshold is set to 0.55. Track age strategy: Set an age threshold of 2 seconds. When the establishment time of a newly established target track is less than the age threshold, the basic decision threshold is reduced by 10% to 20%. For target tracks that exceed the age threshold, the decision threshold remains unchanged. In this embodiment, the basic decision threshold is set to 0.7. Signal-to-noise ratio strategy: First, the average noise power of the entire radar detection area is calculated in real time, and the environmental signal-to-noise ratio is estimated. Then, a lower first threshold and a higher second threshold are preset respectively. In this embodiment, the first threshold is set to 8dB, the second threshold is set to 15dB, and the basic decision threshold is set to 0.7.

[0086] When the ambient signal-to-noise ratio (SNR) drops from a good state: as long as it remains above 15 dB, a higher decision threshold is used, i.e., the decision threshold is set to 0.8. Once the ambient SNR falls below 15 dB, the threshold is immediately lowered. Subsequently, if the SNR fluctuates between 8 dB and 15 dB, the threshold will remain at this higher decision threshold, avoiding frequent switching.

[0087] When the ambient signal-to-noise ratio (SNR) recovers from a poor condition: as long as it remains below 8 dB, a low decision threshold is used, i.e., the decision threshold is set to 0.5. Once the ambient SNR exceeds 8 dB, the threshold is not immediately increased; it must wait until the ambient SNR continues to recover to above 15 dB before the threshold jumps back to a higher value. The intermediate 8 dB to 15 dB region acts as a stable buffer, effectively suppressing oscillations.

[0088] When the environmental signal-to-noise ratio is between 8dB and 15dB, the decision threshold of the previous moment remains unchanged.

[0089] Furthermore, to improve the adaptability of the dynamic decision threshold to complex intersection scenarios, this invention can also employ a multi-strategy weighted fusion approach to determine the final decision threshold. Based on the outputs of three basic strategies—scenario type, track age, and hysteresis comparison—threshold optimization is achieved through adaptive weight allocation. The specific steps are as follows: Each policy is assigned a corresponding dynamic weight w1, w2, w3, which reflects the confidence or importance of the corresponding policy at the current time step. In this embodiment, the scenario type strategy weight w1 is set to a fixed high weight of 0.5 because the strategy reflects prior road design knowledge.

[0090] The track age strategy weight w2 decays exponentially with track age, i.e., w2 = e 年龄 / T Where T is a time constant, and in this embodiment, T = 10 frames.

[0091] The hysteresis comparison strategy weight w3 is related to the signal-to-noise ratio (SNR), and is assigned a higher weight when the SNR is poor. Where k and signal-to-noise ratio mid To adjust the parameters, the weight is close to 1 when the signal-to-noise ratio is poor and close to 0.2 when it is good.

[0092] After obtaining all weights, normalize them so that w1+w2+w3=1.

[0093] The final dynamic decision threshold is generated by weighted summation: Finally, the probability of existence will be determined. Compared with the final judgment threshold, if the probability ultimately exists... If the value is not lower than the judgment threshold, the vehicle is determined to exist; otherwise, the vehicle is determined to not exist.

[0094] S6: Final result output; The final judgment result is compared with the identifier of the target track that produced the result, the lane information, and the final existence probability. The message is encapsulated in JSON format and output to the signal controller or traffic management platform via UDP protocol.

[0095] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of this application. Those skilled in the art may find other optimizations and additional functions in this application. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for detecting vehicle presence at intersections based on multi-dimensional microwave features and intelligent fusion, characterized in that: Includes the following steps: S1: Multi-source data acquisition; Microwave radar echo signals and real-time traffic light status information are collected by a microwave radar sensor array deployed on the traffic lights at intersections and a traffic light status interface, and pre-stored high-precision map information of the intersections is also obtained. S2: Data preprocessing; The microwave radar echo signal acquired in step S1 is preprocessed to obtain a multi-dimensional data cube containing distance, velocity and angle information. The multidimensional data cube is then processed to obtain a target object composed of multiple scattering point traces, and a target trajectory is formed based on the target object. In the process of forming the target track, each target track is associated with a data segment of its corresponding microwave radar echo signal; The data segment includes at least the slow time series signal corresponding to the target trajectory, the horizontal and vertical dual-polarized channel signal, and the broadband echo signal; The target trajectory has a unique ID and contains at least real-time position coordinates and radial velocity estimates; S3: Feature extraction; For each of the target tracks formed in step S2, multidimensional feature extraction is performed, wherein the multidimensional features include at least microDoppler features, polarization features, and motion and geometric features; The extracted multidimensional features are normalized and concatenated into a multidimensional feature vector F; S4: Generate preliminary confidence probabilities; The multidimensional feature vector F obtained in step S3 is input into a pre-trained machine learning classification model, which outputs a continuous value between [0,1] as the initial confidence probability P. S5: Intelligent Decision Making; Includes the following sub-steps: S501: Based on the target trajectory, real-time traffic light status information, and pre-stored high-precision intersection map, the preliminary confidence probability P obtained in step S4 is regularized and corrected to obtain the corrected probability p; S502: Utilizing the continuity of the target trajectory, perform temporal smoothing on the correction probability p of multiple consecutive frames of the same target trajectory to obtain the final existence probability. ; S503: Preset a decision threshold to determine the final existence probability. The result is compared with the judgment threshold, and based on the comparison result, a final judgment result on whether the vehicle exists is generated; S6: Final result output; The final judgment result is compared with the identifier of the target track that produced the result, the lane information, and the final existence probability. Output them together.

2. The intersection vehicle presence detection method based on multi-dimensional microwave features and intelligent fusion as described in claim 1, characterized in that: In step S2, the dot processing includes the following sub-steps: S201: On the multidimensional data cube, a constant false alarm rate (CFAR) detection algorithm is used to identify significant scattering points, wherein the trace of each scattering point includes at least its distance, radial velocity, azimuth angle, and intensity information; S202: Using an aggregation algorithm, the traces of the scattering points that are spatially adjacent and have the same motion state are aggregated to form a target object representing a single physical object; S203: Based on the target object in the current frame, a multi-target tracking algorithm is used to associate it with the trajectory of historical frames and update its state to form a target trajectory; the current estimated position and velocity of the target trajectory are output in real time by the state estimator in the multi-target tracking algorithm.

3. The intersection vehicle presence detection method based on multi-dimensional microwave features and intelligent fusion as described in claim 1, characterized in that: In step S2, the multidimensional data cube also contains complex data of horizontal and vertical polarization channels corresponding to each data unit.

4. The intersection vehicle presence detection method based on multi-dimensional microwave features and intelligent fusion as described in claim 1, characterized in that: In step S3, the extraction of the micro-Doppler features specifically involves: acquiring a slow time-series signal associated with the target trajectory, performing a short-time Fourier transform on it to obtain a micro-Doppler time spectrum, and extracting at least one time-frequency statistical feature value from the micro-Doppler time spectrum as a micro-Doppler feature.

5. The intersection vehicle presence detection method based on multi-dimensional microwave features and intelligent fusion as described in claim 1, characterized in that: In step S3, the extraction of the polarization features specifically involves: acquiring the horizontal polarization channel signal and the vertical polarization channel signal associated with the target track, respectively, calculating the amplitude ratio and phase difference of the two channel signals, and using the amplitude ratio and phase difference to constitute the polarization features.

6. The intersection vehicle presence detection method based on multi-dimensional microwave features and intelligent fusion as described in claim 1, characterized in that: In step S3, the extraction of motion and geometric features specifically involves: From the tracking status of the target trajectory, obtain the radial velocity v of the current target trajectory and the radar cross section calculated based on the microwave radar echo signal power corresponding to the target trajectory; Simultaneously, a broadband echo signal corresponding to the target trajectory is acquired, and a one-dimensional high-resolution range image is obtained by pulse compression processing of the broadband echo signal. From the one-dimensional high-resolution range image, the number of peak points exceeding a preset amplitude threshold is extracted as the number of scattering points feature, and the radial distance span from the first peak point to the last peak point in the range image is extracted as the length feature. The motion and geometric features include at least one or more of the following: the radial velocity v of the target trajectory, the radar cross section, the number of scattering points, and the length feature.

7. The intersection vehicle presence detection method based on multi-dimensional microwave features and intelligent fusion as described in claim 1, characterized in that: In step S3, the multidimensional features also include energy domain features. The specific extraction process is as follows: extract the one-dimensional range image of the zero Doppler channel of the continuous multi-frame series associated with the target track, construct a time series based on the peak amplitude of the one-dimensional range image, and calculate the variance of the time series. The variance of the time series is inversely proportional to the stability of the echo energy.

8. The intersection vehicle presence detection method based on multi-dimensional microwave features and intelligent fusion as described in claim 1, characterized in that: In step S501, the rule-based correction specifically involves: applying predefined rules corresponding to the traffic light phase and spatial area to correct the initial confidence probability P based on the real-time traffic light status information and the position and movement status of the target trajectory in the high-precision map of the intersection, thereby obtaining the corrected probability p. The predefined rules include: When the current traffic light phase is red, determine whether the target trajectory is located within a preset waiting area before the stop line of the lane corresponding to the red light phase, and whether the absolute value of the radial speed of the target trajectory is less than a preset first speed threshold; if it is located, apply a first weighting factor greater than 1 to the initial confidence probability P for positive gain correction; if it is not located, apply a second weighting factor not greater than 1 to the initial confidence probability P for correction, or leave the initial confidence probability P unchanged. When the current traffic light phase is green, determine the movement state of the target trajectory within the green light lane; if its radial velocity is greater than a preset second movement threshold, it is considered to be moving, and a third weighting factor of approximately 1 is applied to the initial confidence probability P for correction or the initial confidence probability P is kept unchanged; if its radial velocity is less than a preset third movement threshold, and the duration of this state exceeds a preset time window, it is considered to be stationary or moving at extremely low speed, and a fourth weighting factor less than 1 is applied to the initial confidence probability P for negative suppression correction. When the current traffic light phase is yellow, based on the target track's current position, speed, and remaining traffic light time, predict whether it can cross the stop line before the red light turns on; if it is predicted that it can, then the green light phase passage rules are used for correction; if it is predicted that it cannot, then the red light phase waiting area rules are used for correction. The correction probability p is limited to no more than 1.

0.

9. The intersection vehicle presence detection method based on multi-dimensional microwave features and intelligent fusion as described in claim 1, characterized in that: In step S502, the time-series smoothing process specifically involves using a Kalman filter algorithm or an exponentially weighted moving average algorithm to smooth the correction probability p of the same target trajectory in multiple processing cycles that are adjacent in time.

10. The intersection vehicle presence detection method based on multi-dimensional microwave features and intelligent fusion as described in claim 1, characterized in that: In step S503, the dynamic decision threshold is dynamically set according to at least one of the following strategies: Scene type strategy: Determine the decision threshold based on the predefined detection scene type of the area where the target track is located; Track age: Set an age threshold. If the establishment time of a newly established track is less than the age threshold, the decision threshold is reduced by 10% to 20%. Hysteresis comparison strategy: A first threshold and a second threshold are preset respectively. Based on the environmental signal-to-noise ratio calculated from the microwave radar echo signal, when the environmental signal-to-noise ratio is lower than the first threshold, the decision threshold is lowered; when the environmental signal-to-noise ratio is higher than the second threshold, the decision threshold is raised; when the environmental signal-to-noise ratio is between the first threshold and the second threshold, the decision threshold of the previous moment remains unchanged.

Images