CN-121811664-B - Road side traffic monitoring method and system based on single millimeter wave radar

CN121811664BCN 121811664 BCN121811664 BCN 121811664BCN-121811664-B

Abstract

The application belongs to the technical field of traffic detection, and particularly relates to a road side traffic monitoring method and system based on a single millimeter wave radar, wherein the road side traffic monitoring method and system comprises a data acquisition module, a millimeter wave point cloud extraction and enhancement module, a traffic flow statistics module, a vehicle speed estimation module and a three-dimensional imaging module; by virtue of the characteristics of all weather millimeter wave signals, strong anti-interference capability, no need of illumination, non-contact sensing and the like, the system can still stably operate in foggy days, nights and complex traffic environments, and the application scene is wider. The influence of shielding among vehicles is effectively relieved by introducing a space-time diversity mechanism, an anchor point type speed measuring method is provided, the speed measuring precision is ensured, the imaging resolution is considered, and the high-precision three-dimensional reconstruction is realized by combining an efficient imaging optimization model, an enhanced representation and a structure guiding strategy. By acquiring the number of vehicles, the running speed of the vehicles and the three-dimensional imaging information of the vehicles, multidimensional road traffic perception data can be formed, so that basic data support is provided for the intelligent traffic system.

Inventors

XU WEI
YANG YANNI
ZOU WEIJUN
LI ZEZHAO
HAN MINGDA
XUE MENG
Cao Yetong
ZHANG GUOMING
HU PENGFEI

Assignees

山东大学

Dates

Publication Date: 20260512
Application Date: 20260311

Claims (9)

1. The road side traffic monitoring method based on the single millimeter wave radar is characterized by comprising the following steps of: S1, road side vehicle data acquisition; s2, millimeter wave point cloud extraction and enhancement; s3, carrying out traffic flow statistics and vehicle speed estimation based on continuous frame point clouds; the traffic flow statistics steps are as follows: The radar main beam direction is set as a detection line, and the radar effective field of view is defined as a detection area, namely, for each frame The extracted point cloud of each vehicle Divided into In the individual lane areas; calculating centroids of corresponding vehicle clusters And assigning lane numbers to each centroid ; Tracking the vehicle track in the continuous frames by adopting a Kalman filter, wherein the centroid of the current frame is matched with the existing track through a Hungary algorithm; When the detection of the intersection of the track and the detection line, recording the corresponding lane, adding 1 to the traffic flow, and obtaining the traffic flow of a certain time period by counting the number of vehicles passing through the detection line in each lane; the vehicle speed estimation steps are as follows: dividing the vehicle head by extracting the points of the vehicle point cloud cluster, wherein the X-axis coordinates of the points are positioned in a preset range in front of the vehicle body, and defining a head quality score: ; Measuring the smoothness of the head point cloud in the Z-axis direction; The regularity of the head of the vehicle in the X-axis direction is measured, and the calculation formula is as follows: ; Wherein: the standard deviation of the head point on the Z-axis coordinate; is the total number of head points; Represent the first X-axis coordinates of the points; an X-axis coordinate representing a current fine candidate position; at the time of obtaining each candidate fine position Then, selecting from the left and right roughing areas The highest position is used as a final anchor point; when the vehicle sequentially passes through the two anchor points, calculating the corresponding frame difference and converting the frame difference into a time interval, wherein the speed of the vehicle is obtained by dividing the distance between the anchor points by the time interval; s4, carrying out vehicle three-dimensional imaging based on the enhanced point cloud; And then taking Gaussian noise as an initial generation state, carrying out partial replacement and constraint on initial noise distribution by utilizing strong reflection high confidence millimeter wave points, constructing a calibrated initial state, enhancing the consistency with a real vehicle structure, inputting the condition characteristics into an imaging network, and realizing three-dimensional imaging output of the vehicle.
2. The road side traffic monitoring method based on single millimeter wave radar according to claim 1, wherein the step S2 of millimeter wave point cloud extraction and enhancement comprises the following specific steps: s21, reconstructing single frame data into a slow time-fast time two-dimensional matrix, segmenting the original frame data according to a fast time dimension to obtain a plurality of sub-segment data, and respectively carrying out distance and Doppler processing on the complete frame and each sub-segment to form a multi-scale distance-speed diagram; s22, calculating a self-adaptive soft threshold value based on a complete frame result, carrying out target screening on each scale result, and extracting an effective target point on a distance-speed diagram; S23, estimating azimuth angle and pitch angle of a target point by utilizing multi-antenna Bartlett wave beam forming, and converting the azimuth angle and pitch angle into three-dimensional space coordinates by combining distance information; s24, carrying out lane partitioning and density clustering on the generated three-dimensional point cloud to obtain an enhanced point cloud corresponding to each vehicle.
3. The method for road side traffic monitoring based on single millimeter wave radar according to claim 2, wherein in step S21, the original signal is subjected to segmentation processing, and an intra-chirp segmentation method is used, or/and, an inter-chirp segmentation method is used: Using intra-chirp segmentation, a single frame radar signal is set to a two-dimensional matrix x (n, m), where Representing a slow time dimension chirp index, Representing fast-time dimension sample point index along fast-time dimension Dividing a single frame signal into A sub-segment, which is mathematically expressed as: ; Wherein, the The segmentation mode corresponds to a coarse-scale sampling result of the distance dimension; in inter-chirp segmentation, along the slow time dimension Dividing a single frame signal into A sub-segment, which is mathematically expressed as: ; Wherein, the The segmentation mode corresponds to a coarse-scale sampling result of a speed dimension; After the segmentation is completed, a signal set containing the original complete frame signal and the segmented sub-segment signal Wherein And (2) and A Doppler-FFT operation is then performed on each set of signals in the set, generating a corresponding range-Doppler RDM matrix: ; Wherein: 。
4. the method for road side traffic monitoring based on single millimeter wave radar according to claim 3, wherein step S22 extracts a target area of the vehicle in the generated RDM matrix, specifically: First, in the original RDM graph without segmentation Obtaining a core region by adopting a conservative threshold; Subsequently, by fusion function Generating a joint quality evaluation index for adjusting the adaptive coefficient: ; Wherein: And For avoiding over-selection and under-selection, respectively; SNR characterizes the prominence of the strongest target echo relative to background noise; PV is the number of single cluster points after DBSCAN clustering and is used for describing the spatial scale of a target; SC is the ratio of the number of points to the area of the bounding box, and is used for measuring the space density; adopts a Sum-Product fusion strategy: ; Wherein, the Is a balance coefficient; finally according to Extracting target areas in RDM of different scales, And Corresponding to the mean and standard deviation of the RDM of different scales respectively.
5. The single millimeter wave radar-based roadside traffic monitoring method according to claim 1, wherein in step S3, in case of inter-vehicle occlusion, the historical track is based Using Kalman filtering device pair shielded vehicle Continuously predicting the missing centroid of (2) to obtain a predicted centroid When new effective vehicle centroid When detected, the prediction process is stopped and the predicted centroid and the new centroid are sequentially connected into the historical track, thereby restoring the complete track.
6. The single millimeter wave radar-based roadside traffic monitoring method according to claim 1, wherein in step S3, noise point cloud processing: The first case is that the historical track may be derived from a continuous noise point cloud, rather than a real vehicle point cloud, for a predicted centroid Detecting whether there is a preceding vehicle The method comprises the following steps: and the space area occupied by the vehicle point cloud cluster With radar position Connection to predicted centroid There is a spatial intersection: ; Wherein, the Representing a set of all vehicle centroids; Representing a vehicle A lane positioned in front of the vehicle to be compensated; Representing radar position A line connecting the predicted centroid; The space area occupied by the point cloud cluster is represented, and the predicted mass center is judged to meet the physical shielding relation and is reserved only when the conditions are met, otherwise, compensation is stopped to avoid forming false tracks; The second case is that the newly detected centroid Also from the noise point cloud, the distance between the new centroid and the kalman predicted centroid is compared: ; if the distance is smaller than the history track Average width of midpoint cloud And judging the new centroid as a valid observation and accessing the track, and discarding the centroid otherwise.
7. The single millimeter wave radar-based roadside traffic monitoring method according to claim 1, wherein step S4 performs vehicle three-dimensional imaging based on the enhanced point cloud, comprising model training and model reasoning: The model training steps are as follows: from Gaussian noise distribution Point cloud distribution to real vehicles Is a linear generation path of: ; Wherein, the Representing a sampling time step; the linear path is defined by the velocity field Driving to make the intermediate state Realizing speed field prediction by adopting a U-Net structure, and carrying out regression approximation on a real speed field: ; constructing an enhancement condition representation by adopting a bird's eye view projection and distance map projection splicing fusion mode: ; the network optimization objective is to minimize the distance between the predicted speed field and the true speed field Distance: ; Wherein, the Representing conditional guidance information constructed from a radar point cloud; The enhanced condition input U-Net is embedded in multiple layers, and the condition features are encoded into multiple layers of independent inputs in order to avoid the attenuation of the condition information in the deep propagation process Injecting U-Net layers respectively to realize multi-scale structure constraint, and simultaneously, time steps Introducing conditional expressions, and realizing time-aware embedding through a scale modulation and offset modulation mechanism: ; And Respectively represent by parameters A characterized scaling function and an offset function for time-dependent variables For conditional characteristics Performing linear modulation; in the reasoning stage, selecting intensity values located at high confidence points to construct a reference matrix And with initial noise Fusion: ; Wherein, the Is a matrix of binary masks that are to be applied, For a full 1 matrix, the velocity field is calibrated at the same time: ; the method is used for controlling the guiding strength of the high confidence point in the initial reasoning stage; From the initial noise Starting from discrete steps Gradually updating: ; Representing the parameters by A characterized velocity field function for use in the condition information To give state variables under the constraint of (2) At the time of The instantaneous change rate of evolution along the generation path is output as a change vector of state variables for determining the point cloud state in time step The update direction and magnitude within; and (5) obtaining the reconstructed vehicle three-dimensional structure through iteration.
8. The single millimeter wave radar-based roadside traffic monitoring method of claim 7 wherein a reconstruction regression loss term is introduced based on a speed regression loss: ; The deduction can be obtained: ; representing the weight coefficient of the reconstruction loss as a positive number for regulating the reconstruction loss The influence degree in the overall loss function is used for controlling the contribution proportion of the output point cloud reconstruction loss to the model optimization process; representing cloud state of target time point predicted by model, namely current state In the velocity field Under the action of (a) deducing the obtained reconstruction point cloud result for being used for matching with the real point cloud And calculating a reconstruction error.
9. The road side traffic monitoring system based on the single millimeter wave radar is characterized by comprising a data acquisition module, a millimeter wave point cloud extraction and enhancement module, a traffic flow statistics module, a vehicle speed estimation module and a three-dimensional imaging module; the data acquisition module is used for fixedly mounting the millimeter wave radar on the road side, so that the main beam direction of the radar is perpendicular to the running direction of the vehicle and covers a multi-lane road area; The millimeter wave point cloud extraction and enhancement module comprises a first step of reconstructing single frame data into a slow time-fast time two-dimensional matrix, segmenting the original frame data according to a fast time dimension to obtain a plurality of sub-segment data, a second step of respectively carrying out distance and Doppler processing on a complete frame and each sub-segment to form a multi-scale distance-speed graph, a third step of calculating a self-adaptive soft threshold value based on the complete frame result, carrying out target screening on each scale result, extracting an effective target point on the distance-speed graph, a fourth step of estimating the azimuth angle and the pitch angle of the target point by utilizing multi-antenna Bartlett wave beam formation on the basis, and converting the azimuth angle and the pitch angle into three-dimensional space coordinates by combining distance information; The traffic flow statistics module is used for converting captured original signals of continuous frames of vehicles into millimeter wave point cloud data of the continuous frames; based on the continuous frame point clouds, setting a detection area and a detection line in the radar view field range, carrying out lane division and clustering on each frame of point clouds, extracting each vehicle point cloud cluster, calculating the corresponding centroid position, and simultaneously giving lane numbers; the method comprises the steps of establishing a vehicle state model based on Kalman filtering, predicting and updating a continuous frame centroid, completing matching between the current frame centroid and a historical track by combining a Hungary algorithm, initializing an unmatched target into a new track, realizing continuous tracking of vehicles, judging whether each vehicle track passes through a preset detection line after the track is formed, counting and accumulating according to a belonging lane if the track passes through, performing track prediction compensation by using Kalman filtering when the centroid is temporarily lost due to shielding, performing validity matching and updating on the newly-appearing centroid in a subsequent frame, and finally counting the number of vehicles passing through the detection line in a set time window to obtain a vehicle flow result of a corresponding time period; The vehicle speed estimation module is used for calculating the vehicle running speed through space displacement and time difference on the basis of obtaining enhanced continuous frame vehicle point cloud data, firstly determining two rough anchoring areas on two sides of a road, then extracting the vehicle head point cloud of the vehicle in each rough area as a candidate anchor point, carrying out matching calculation on the vehicle head point cloud and a plurality of fine positions preset in the rough area, and jointly determining the quality score by the discrete degree of the vehicle head point in the vertical direction and the fitting compactness along the running direction, selecting the fine position with the highest score as a final space anchor point, then recording the time difference when the same vehicle head sequentially passes through the two anchor points, combining the known space distance between the two anchor points, and obtaining the vehicle average running speed according to the ratio of the distance to the time; The three-dimensional imaging module is used for converting input point clouds into bird's eye view and distance graph representation respectively, splicing and fusing the bird's eye view and distance graph representation to form unified condition features, taking Gaussian noise as an initial generation state, carrying out partial replacement and constraint on initial noise distribution by utilizing strong reflection high confidence millimeter wave points, constructing a calibrated initial state to enhance the consistency with a real vehicle structure, inputting the condition features into an imaging network, combining time step parameter modulation, updating the point cloud state through discrete time step iteration, gradually approaching the real vehicle three-dimensional structure under condition guidance, and finally carrying out back projection on the generated result to the three-dimensional space to output a vehicle three-dimensional imaging point cloud result.

Description

Road side traffic monitoring method and system based on single millimeter wave radar Technical Field The application belongs to the technical field of traffic detection, and particularly relates to a road side traffic monitoring method and system based on a single millimeter wave radar. Background Modern intelligent traffic systems are moving from passive data acquisition to comprehensive traffic digital twinning, requiring real-time perception of vehicle semantics, kinematics and geometric features. The traditional aerial view angle monitoring relies on a special portal frame or overpass structure, the visual field is not blocked, but the construction and maintenance cost is high, the deployment is complex, the large-scale popularization is difficult, the unmanned aerial vehicle scheme deployment is flexible, the unmanned aerial vehicle scheme deployment is limited by endurance and stability, and long-term continuous operation is difficult to support. In contrast, the road side monitoring mode based on the existing infrastructure such as the street lamp post and the traffic sign is more expandable, and has the potential of constructing a distributed continuous perception network. However, the existing road side system mostly adopts a multi-sensor splicing architecture, different functions are respectively realized by different types of sensors, so that the system has high cost, complex structure and difficult maintenance, and meanwhile, the visual sensor has obviously degraded performance under severe environments such as rain, fog, snow and the like, and the all-weather stable operation is difficult to ensure. The millimeter wave radar has all-weather working and space sensing capabilities and has the potential of being used as a unified sensing carrier. However, when traffic flow statistics, speed estimation and vehicle three-dimensional imaging are simultaneously realized on a single millimeter wave radar platform, single-function optimization strategies are difficult to directly superpose, and different tasks have inherent conflicts in signal parameter configuration and system resource allocation (1) enhancing the visibility of an occlusion target depends on multi-frame superposition, reducing time resolution and affecting high-speed target observation, (2) enlarging a non-fuzzy speed range, needing to adjust modulation parameters, possibly sacrificing distance resolution and affecting fine imaging, and (3) the computational complexity required by high-quality vehicle-level imaging is contradictory with the real-time low-delay processing requirement of a road side. Therefore, the prior art is difficult to realize unified and stable output of the multidimensional traffic information under the condition of low cost. Disclosure of Invention Based on the problems, the invention provides a road side traffic monitoring method and a road side traffic monitoring system for a single millimeter wave radar, which can realize traffic flow statistics, vehicle speed estimation and vehicle three-dimensional imaging under a unified framework in a cooperative manner, and provide a low-cost integrated solution for robust traffic monitoring under dark and extreme environmental conditions. The technical proposal is as follows: a road side traffic monitoring method based on a single millimeter wave radar comprises the following steps: S1, road side vehicle data acquisition; s2, millimeter wave point cloud extraction and enhancement; s3, carrying out traffic flow statistics and vehicle speed estimation based on continuous frame point clouds; s4, carrying out vehicle three-dimensional imaging based on the enhanced point cloud; And then taking Gaussian noise as an initial generation state, carrying out partial replacement and constraint on initial noise distribution by utilizing strong reflection high confidence millimeter wave points, constructing a calibrated initial state, enhancing the consistency with a real vehicle structure, inputting the condition characteristics into an imaging network, and realizing three-dimensional imaging output of the vehicle. Preferably, the step S2 of millimeter wave point cloud extraction and enhancement specifically comprises the following steps: s21, reconstructing single frame data into a slow time-fast time two-dimensional matrix, segmenting the original frame data according to a fast time dimension to obtain a plurality of sub-segment data, and respectively carrying out distance and Doppler processing on the complete frame and each sub-segment to form a multi-scale distance-speed diagram; s22, calculating a self-adaptive soft threshold value based on a complete frame result, carrying out target screening on each scale result, and extracting an effective target point on a distance-speed diagram; S23, estimating azimuth angle and pitch angle of a target point by utilizing multi-antenna Bartlett wave beam forming, and converting the azimuth angle and pitch angle into three-dimensional space