Single millimeter wave radar based roadside traffic monitoring method and system

Abstract

The application belongs to the technical field of traffic detection, and particularly relates to a roadside traffic monitoring method and system based on a single millimeter wave radar, which comprises a data acquisition module, a millimeter wave point cloud extraction and enhancement module, a vehicle flow statistical module, a vehicle speed estimation module and a three-dimensional imaging module; thanks to the characteristics of the millimeter wave signal, such as all-weather, strong anti-interference ability, no need for light and non-contact sensing, the system can still operate stably under foggy weather, night and complex traffic environment, and is more widely applicable. The introduction of the spatiotemporal diversity mechanism effectively alleviates the influence of occlusion between vehicles, an anchor point speed measurement method is proposed, the imaging resolution is taken into account while ensuring the speed measurement accuracy, and the high-precision three-dimensional reconstruction is realized by combining the efficient imaging optimization model and the enhanced representation and structure guiding strategy. By obtaining the number of vehicles, the vehicle driving speed and the three-dimensional imaging information of the vehicles, multi-dimensional road traffic perception data can be formed, thereby providing basic data support for the intelligent traffic system.

Description

Technical Field

[0001] This application belongs to the field of traffic detection technology, specifically relating to a roadside traffic monitoring method and system based on a single millimeter-wave radar. Background Technology

[0002] Modern intelligent transportation systems are moving from passive data collection to comprehensive digital twins of transportation, requiring real-time perception of the semantic, kinematic, and geometric characteristics of vehicles. Traditional bird's-eye view monitoring relies on dedicated gantry structures or overpasses, offering unobstructed views but incurring high construction and maintenance costs and complex deployments, hindering large-scale deployment. While drone solutions offer flexible deployment, their limited range and stability make long-term continuous operation difficult. In contrast, roadside monitoring based on existing infrastructure such as streetlights and traffic signs offers greater scalability and the potential to build distributed, continuous sensing networks. However, existing roadside systems often employ multi-sensor architectures, with different functions implemented by different types of sensors, resulting in high system costs, complex structures, and difficult maintenance. Furthermore, visual sensors experience significant performance degradation in harsh environments such as rain, fog, and snow, making it difficult to guarantee stable all-weather operation. Millimeter-wave radar, with its all-weather operation and spatial perception capabilities, has the potential to serve as a unified sensing platform. However, when traffic flow statistics, speed estimation, and vehicle 3D imaging are simultaneously achieved on a single millimeter-wave radar platform, single-function optimization strategies are difficult to directly superimpose. Different tasks have inherent conflicts in terms of signal parameter configuration and system resource allocation: (1) Enhancing the visibility of occluded targets depends on multi-frame superposition, which reduces temporal resolution and affects the observation of high-speed targets; (2) Expanding the unambiguous speed range requires adjusting modulation parameters, which may sacrifice range resolution and affect fine imaging; (3) The computational complexity required for high-quality vehicle-level imaging contradicts the requirements of real-time low-latency processing on the roadside. Therefore, existing technologies are unable to achieve unified and stable output of multi-dimensional traffic information under low-cost conditions. Summary of the Invention

[0003] To address the aforementioned problems, this invention proposes a roadside traffic monitoring method and system using a single millimeter-wave radar. This system can collaboratively achieve traffic flow statistics, vehicle speed estimation, and 3D vehicle imaging within a unified architecture, providing a low-cost, integrated solution for robust traffic monitoring under dark and extreme environmental conditions. The technical solution is as follows:

[0004] A roadside traffic monitoring method based on a single millimeter-wave radar includes the following steps:

[0005] S1. Roadside vehicle data collection;

[0006] S2. Millimeter-wave point cloud extraction and enhancement;

[0007] S3. Traffic flow statistics and vehicle speed estimation based on continuous frame point clouds;

[0008] S4. Vehicle 3D Imaging Based on Enhanced Point Cloud;

[0009] The input point cloud is converted into a bird's-eye view and a distance map representation, which are then stitched together to form a unified conditional feature. Subsequently, Gaussian noise is used as the initial generation state, and the initial noise distribution is partially replaced and constrained by high-confidence millimeter-wave points with strong reflection to construct a calibrated initial state to enhance consistency with the real vehicle structure. The conditional feature is then input into the imaging network to achieve 3D vehicle imaging output.

[0010] Preferably, step S2, millimeter-wave point cloud extraction and enhancement, includes the following specific steps:

[0011] S21. Reconstruct the single-frame data into a slow-time–fast-time two-dimensional matrix, segment the original frame data according to the fast-time dimension to obtain multiple sub-segments; perform distance and Doppler processing on the complete frame and each sub-segment to form a multi-scale distance-velocity map.

[0012] S22. Calculate an adaptive soft threshold based on the complete frame results, perform target filtering on the results at each scale, and extract the effective target points on the distance-velocity map;

[0013] S23. Use multi-antenna Bartlett beamforming to estimate the azimuth and elevation angles of the target point, and combine them with range information to convert them into three-dimensional spatial coordinates;

[0014] S24. Perform lane partitioning and density clustering on the generated 3D point cloud to obtain the enhanced point cloud corresponding to each vehicle.

[0015] Preferably, in step S21, the original signal is segmented using intra-chirp segmentation and / or inter-chirp segmentation.

[0016] Intra-chirp segmentation: Let a single-frame radar signal be a two-dimensional matrix x(n,m), where This represents the slow-time dimension chirp index. Represents the index of the sampling point in the fast time dimension, along the fast time dimension. Divide a single frame signal into Each sub-segment, mathematically expressed as:

[0017] ;

[0018] in, This segmentation method corresponds to the coarse-scale sampling results of the distance dimension.

[0019] In the inter-chirp partitioning approach, along the slow time dimension Divide a single frame signal into Each sub-segment, mathematically expressed as:

[0020] ;

[0021] in, This segmentation method corresponds to the coarse-scale sampling results of the velocity dimension;

[0022] After segmentation, the signal set includes the original complete frame signal and the segmented sub-segment signals. ,in ,and Then, a Doppler-FFT operation is performed on each group of signals in the set to generate the corresponding range-Doppler RDM matrix:

[0023] ;

[0024] in:

[0025] .

[0026] Preferably, in step S22, the target region of the vehicle is extracted from the generated RDM matrix, specifically as follows:

[0027] First, in the original, unsegmented RDM graph The core region is obtained by using a conservative threshold.

[0028] Subsequently, through the fusion function Generate joint quality evaluation indicators to adjust the adaptive coefficients:

[0029] ;

[0030] in:

[0031] and These are used to avoid over-selection and under-selection, respectively.

[0032] SNR characterizes the prominence of the strongest target echo relative to background noise;

[0033] PV is the number of points in a single cluster after DBSCAN clustering, used to characterize the spatial scale of the target.

[0034] SC is the ratio of the number of points to the area of ​​its bounding box, used to measure the density of space.

[0035] Adopting a Sum-Product fusion strategy:

[0036] ;

[0037] in, This is the balance coefficient;

[0038] Ultimately based on Extract target regions from RDM at different scales. and These correspond to the mean and standard deviation of RDM at different scales.

[0039] Preferably, in step S3, the traffic flow statistics steps are as follows:

[0040] First, the radar main beam direction is set as the detection line, and the radar effective field of view is defined as the detection area: for each frame The point cloud of each vehicle will be extracted. Divided into In each lane area;

[0041] Then the centroid of the corresponding vehicle cluster is calculated. And assign lane numbers to each centroid. ;

[0042] Next, a Kalman filter is used to track vehicle trajectories in consecutive frames. The centroid of the current frame is matched with existing trajectories using the Hungarian algorithm. Unmatched centroids initialize new trajectories based on their geometric positions and lane information, thereby ensuring continuous tracking of newly appearing vehicles.

[0043] For each trajectory, check whether it intersects with the detection line: when the trajectory intersects with the detection line, record the corresponding lane and increment the traffic flow by 1. By counting the number of vehicles passing through the detection line in each lane within a certain time period, obtain the traffic flow for that time period.

[0044] The steps for estimating vehicle speed are as follows:

[0045] First, the vehicle head is segmented by extracting points in the vehicle point cloud cluster whose X-axis coordinates are located within a set range in front of the vehicle body.

[0046] Subsequently, the head quality score was defined:

[0047] ;

[0048] Measure the smoothness of the head point cloud in the Z-axis direction; The regularity of the vehicle's front end along the X-axis is measured; the formula for its calculation is:

[0049] ;

[0050] in:

[0051] This represents the standard deviation of the head point in the Z-axis coordinate system.

[0052] This represents the total number of head points;

[0053] Indicates the first The X-axis coordinates of each point;

[0054] The X-axis coordinates of the current refined candidate position;

[0055] In obtaining the fine location of each candidate Then, select from the left and right coarse selection areas respectively. The highest position is used as the final anchor point;

[0056] Finally, as the vehicle passes through two anchor points in sequence, the corresponding frame difference is calculated and converted into a time interval. The vehicle speed is obtained by dividing the distance between the anchor points by this time interval.

[0057] Preferably, in step S3, when there is occlusion between vehicles, based on historical trajectories... Using a Kalman filter to detect the obscured vehicle The missing centroids are continuously predicted to obtain the predicted centroids. When the new effective vehicle center of gravity When detected, the prediction process is stopped, and the predicted centroid and the new centroid are sequentially connected to the historical trajectory to recover the complete trajectory.

[0058] Preferably, in step S3, noise point cloud processing is performed.

[0059] The first scenario is that the historical trajectory may originate from continuous noisy point clouds rather than actual vehicle point clouds, which affects the prediction of the centroid. Detect whether there is a vehicle ahead. ,satisfy:

[0060] Furthermore, the spatial area occupied by the vehicle point cloud cluster With radar position The line connecting to the predicted centroid Existence of spatial intersection: ;

[0061] in, This represents the set of the centroids of all vehicles. Indicates vehicle Located in the lane in front of the vehicle awaiting compensation; Indicates radar location The line connecting the predicted centroid; The predicted centroid is determined to satisfy the physical occupancy relationship and is retained only when the above conditions are met; otherwise, compensation is terminated to avoid the formation of false trajectories.

[0062] The second scenario is: newly detected centroids. Also derived from noisy point clouds, compare the distance between the new centroid and the Kalman-predicted centroid:

[0063] ;

[0064] If this distance is smaller than the historical trajectory Average width of midpoint cloud If the observation is valid, the new centroid is considered a valid observation and is included in the trajectory; otherwise, the centroid is discarded.

[0065] Preferably, step S4 involves performing 3D vehicle imaging based on the enhanced point cloud, including model training and model inference:

[0066] The model training steps are as follows:

[0067] From Gaussian noise distribution To the actual vehicle point cloud distribution Linear generation path:

[0068] ;

[0069] in, Indicates the sampling time step;

[0070] Linear path by velocity field Drive, to intermediate state The flow is directed towards the target distribution; a U-Net structure is used to predict the velocity field, and a regression approximation is performed on the true velocity field.

[0071] ;

[0072] An enhanced conditional representation is constructed by stitching together bird's-eye view projection and distance map projection:

[0073] ;

[0074] The network optimization objective is to minimize the difference between the predicted velocity field and the true velocity field. distance:

[0075] ;

[0076] in, This represents conditional guidance information constructed from radar point clouds;

[0077] Subsequently, the enhanced conditional input to the U-Net is embedded in multiple layers. To avoid the attenuation of conditional information during deep propagation, the conditional features are encoded as multiple independent inputs. This is injected into each layer of U-Net to achieve multi-scale structural constraints; simultaneously, the time steps are... By introducing conditional representations, time-aware embedding is achieved through scale modulation and offset modulation mechanisms:

[0078] This enhances the model's ability to perceive the location of the generated path;

[0079] and They represent the parameters respectively. The scaling and offset functions are represented and used to scale the time variable. Conditional features Perform linear modulation;

[0080] During the inference phase, a reference matrix is ​​constructed by selecting strength values ​​located at high confidence points. and with initial noise Fusion:

[0081] ;

[0082] in, It is a binary mask matrix. The matrix is ​​composed of all 1s, and the velocity field is calibrated simultaneously.

[0083] ;

[0084] This is used to control the guiding strength of high confidence points in the early stages of inference, and to gradually weaken them in the later stages;

[0085] From initial noise Set off, according to the distance from the walk Gradually updated:

[0086] ;

[0087] Indicates by parameters The velocity field function is characterized and used in the context of conditional information. State variables are given under constraints In time The instantaneous rate of change of the point cloud along the generation path is output as a vector of changes in the state variables, used to determine the point cloud state at the time step. The direction and magnitude of updates within;

[0088] The reconstructed three-dimensional structure of the vehicle is obtained through a finite number of iterations.

[0089] Preferably, a reconstruction regression loss term is introduced based on the velocity regression loss:

[0090] ;

[0091] Through derivation, we can obtain:

[0092] ;

[0093] : Represents the reconstruction loss weighting coefficient, which is a positive number used to adjust the reconstruction loss. The degree of influence in the overall loss function is used to control the proportion of the output point cloud reconstruction loss contributing to the model optimization process;

[0094] : Represents the point cloud state at the target time predicted by the model, i.e., the state at the current time. In the velocity field The reconstructed point cloud results obtained under the influence of the function are used to compare with the real point cloud. Calculate the reconstruction error.

[0095] A roadside traffic monitoring system based on a single millimeter-wave radar includes 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.

[0096] Data acquisition module: The millimeter-wave radar is fixedly installed on the side of the road, so that the direction of the radar's main beam is basically perpendicular to the direction of vehicle travel, covering a multi-lane road area; the radar periodically transmits frequency-modulated continuous wave signals and receives the echoes reflected from the vehicle surface and structure to form a continuous multi-frame raw intermediate frequency signal data;

[0097] The millimeter-wave point cloud extraction and enhancement module first reconstructs single-frame data into a slow-time–fast-time two-dimensional matrix. The original frame data is then segmented according to the fast-time dimension to obtain multiple sub-segments. Next, range and Doppler processing is performed on the complete frame and each sub-segment to generate a multi-scale range-velocity map. Then, an adaptive soft threshold is calculated based on the complete frame results, and target filtering is performed on the results at each scale to extract valid target points on the range-velocity map. Based on this, the azimuth and elevation angles of the target points are estimated using a multi-antenna Bartlett beamforming system, and combined with range information, converted into three-dimensional spatial coordinates. Finally, lane partitioning and density clustering are performed on the generated three-dimensional point cloud to obtain the enhanced point cloud corresponding to each vehicle.

[0098] The traffic flow statistics module converts the captured continuous frame raw signals of vehicles into continuous frame millimeter-wave point cloud data. Based on the continuous frame point cloud, a detection area and detection line are set within the radar field of view. Lane division and clustering are performed on each frame point cloud. The point cloud clusters of each vehicle are extracted and their corresponding centroid positions are calculated, and lane numbers are assigned. Subsequently, a vehicle state model is established based on Kalman filtering to predict and update the centroids of continuous frames. The Hungarian algorithm is used to match the centroids of the current frame with historical trajectories. Unmatched targets are initialized as new trajectories to achieve continuous vehicle tracking. After the trajectory is formed, it is determined whether each vehicle trajectory crosses the preset detection line. If a crossing occurs, the count is accumulated according to the lane to which it belongs. When occlusion causes a temporary loss of centroids, Kalman filtering is used for trajectory prediction compensation. In subsequent frames, the re-emerging centroids are matched and updated for effectiveness. Finally, the number of vehicles that pass through the detection line within the set time window is counted to obtain the traffic flow results for the corresponding time period.

[0099] Vehicle speed estimation module: Based on the enhanced continuous frame vehicle point cloud data, the vehicle speed is calculated through spatial displacement and time difference. First, two coarse anchoring regions are determined on both sides of the road. Then, the vehicle front point cloud is extracted as a candidate anchor point in each coarse region and matched with multiple preset fine positions in the coarse region to calculate a quality score. The score is determined by the dispersion of the vehicle front point in the vertical direction and the tightness of the fit along the driving direction. The fine position with the highest score is selected as the final spatial anchor point. Then, the time difference when the front of the same vehicle passes through two anchor points in sequence is recorded, and combined with the known spatial distance between the two anchor points, the average vehicle speed is obtained according to the ratio of distance to time.

[0100] The 3D imaging module converts the input point cloud into a bird's-eye view and a distance map representation, respectively, and then stitches and fuses them to form a unified conditional feature. Subsequently, Gaussian noise is used as the initial generation state, and strong reflective high-confidence millimeter-wave points are used to partially replace and constrain the initial noise distribution to construct a calibrated initial state, thereby enhancing the consistency with the real vehicle structure. The conditional feature is input into the imaging network, and combined with time-step parameter modulation, the point cloud state is updated iteratively through discrete time steps, gradually approximating the real vehicle 3D structure under conditional guidance. Finally, the generated result is back-projected into 3D space to output the vehicle 3D imaging point cloud result.

[0101] Compared with the prior art, the beneficial effects of this application are as follows:

[0102] This application constructs a system based on a single low-cost commercial millimeter-wave radar. This system integrates multiple functions, including traffic flow statistics, vehicle speed estimation, and 3D vehicle imaging, without requiring additional sensors or complex infrastructure. This significantly reduces system construction and maintenance costs, and improves deployment flexibility and engineering feasibility. Furthermore, leveraging the all-weather, strong anti-interference capabilities, light-free operation, and non-contact sensing characteristics of millimeter-wave signals, the system can operate stably in foggy, nighttime, and complex traffic environments, making it applicable to a wider range of scenarios. Addressing the problems of severe multi-vehicle occlusion and the mutual constraints between statistical and imaging tasks in existing technologies, this patent effectively mitigates the impact of vehicle occlusion by introducing a spatiotemporal diversity mechanism. It proposes an anchor-point speed measurement method that ensures both speed measurement accuracy and imaging resolution. Combined with an efficient imaging optimization model and enhanced representation and structure-guided strategies, it achieves high-precision 3D reconstruction. Attached Figure Description

[0103] Figure 1 A flowchart of a roadside traffic monitoring system based on a single millimeter-wave radar;

[0104] Figure 2 This illustrates the split-aggregation methods for intra-chirp and inter-chirp.

[0105] Figure 3 Illustration of traffic flow counting obstruction compensation;

[0106] Figure 4 For vehicle speed estimation based on a coarse-to-fine vehicle front-end alignment method;

[0107] Figure 5 This is a vehicle imaging model based on a conditionally modified flow model;

[0108] Figure 6 A comparison chart showing the accuracy of traffic flow statistics under different lane obstruction conditions;

[0109] Figure 7 A comparison chart of the accuracy of traffic flow statistics under different environmental conditions;

[0110] Figure 8 A comparison chart of the average error in vehicle speed estimation under different lane obstruction conditions;

[0111] Figure 9 A comparison chart of the average error in vehicle speed estimation under different environmental conditions;

[0112] Figure 10 Comparison chart of average speed estimation error at different speed levels;

[0113] Figure 11 A comparison of quantitative vehicle imaging results with existing methods. Detailed Implementation

[0114] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0115] To address the challenge of balancing observation continuity, speed measurement accuracy, and imaging quality in multi-task traffic perception in roadside environments, this invention proposes a roadside traffic perception method based on a single millimeter-wave radar. First, a millimeter-wave radar is deployed on the side of the road to capture raw echo signals from vehicles. Then, the raw signals are converted into point cloud data. Based on this point cloud data, we perform traffic flow statistics, vehicle speed estimation, and vehicle imaging tasks.

[0116] A roadside traffic monitoring system based on a single millimeter-wave radar includes 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.

[0117] Data acquisition module: The millimeter-wave radar is fixedly installed on the side of the road, so that the direction of the radar's main beam is basically perpendicular to the direction of vehicle travel, covering a multi-lane road area; the radar periodically transmits frequency-modulated continuous wave signals and receives the echoes reflected from the vehicle surface and structure, forming a continuous multi-frame raw intermediate frequency signal data.

[0118] The millimeter-wave point cloud extraction and enhancement module first reconstructs single-frame data into a slow-time–fast-time two-dimensional matrix, then segments the original frame data according to the fast-time dimension to obtain multiple sub-segments. Next, it performs range and Doppler processing on the complete frame and each sub-segment to form a multi-scale range-velocity map. Then, it calculates an adaptive soft threshold based on the complete frame results, filters targets at each scale, and extracts valid target points from the range-velocity map. Based on this, it uses a multi-antenna Bartlett beamforming system to estimate the azimuth and elevation angles of the target points, combining this with range information to convert them into three-dimensional spatial coordinates. Finally, it performs lane partitioning and density clustering on the generated three-dimensional point cloud to obtain the enhanced point cloud corresponding to each vehicle.

[0119] The traffic flow statistics module converts the captured continuous frame raw signals of vehicles into continuous frame millimeter-wave point cloud data. Based on the continuous frame point cloud, a detection area and detection line are set within the radar field of view. Lane division and clustering are performed on each frame point cloud, extracting each vehicle point cloud cluster and calculating the corresponding centroid position, while assigning lane numbers. Subsequently, a vehicle state model is established based on Kalman filtering to predict and update the centroid of continuous frames. The Hungarian algorithm is used to match the current frame centroid with historical trajectories. Unmatched targets are initialized as new trajectories to achieve continuous vehicle tracking. After the trajectory is formed, it is determined whether each vehicle trajectory crosses the preset detection line. If a crossing occurs, the count is accumulated according to the lane to which it belongs. When occlusion causes a temporary loss of centroid, Kalman filtering is used for trajectory prediction compensation, and the re-emerging centroids are effectively matched and updated in subsequent frames. Finally, the number of vehicles that pass through the detection line within the set time window is counted to obtain the traffic flow result for the corresponding time period.

[0120] Vehicle speed estimation module: Based on the enhanced continuous frame vehicle point cloud data, the vehicle speed is calculated through spatial displacement and time difference. First, two coarse anchoring regions are determined on both sides of the road. Then, the vehicle front point cloud is extracted as a candidate anchor point in each coarse region and matched with multiple preset fine positions in the coarse region to calculate a quality score. The score is determined by the dispersion of the vehicle front point in the vertical direction and the tightness of the fit along the driving direction. The fine position with the highest score is selected as the final spatial anchor point. Then, the time difference when the front of the same vehicle passes through two anchor points in sequence is recorded, and combined with the known spatial distance between the two anchor points, the average vehicle speed is obtained according to the ratio of distance to time.

[0121] The 3D imaging module converts the input point cloud into a bird's-eye view and a distance map representation, respectively, and then stitches and fuses them to form a unified conditional feature. Subsequently, Gaussian noise is used as the initial generation state, and strong reflective high-confidence millimeter-wave points are used to partially replace and constrain the initial noise distribution to construct a calibrated initial state, thereby enhancing the consistency with the real vehicle structure. The conditional feature is input into the imaging network, and combined with time-step parameter modulation, the point cloud state is updated iteratively through discrete time steps, gradually approximating the real vehicle 3D structure under conditional guidance. Finally, the generated result is back-projected into 3D space to output a high-resolution vehicle 3D imaging point cloud result.

[0122] The vehicle quantity, vehicle speed, and vehicle 3D imaging information acquired by this system together constitute road traffic monitoring data, which can be used for traffic flow analysis, road congestion monitoring, vehicle type identification, and traffic operation status assessment, providing basic perception data support for intelligent transportation systems.

[0123] Figure 1This paper demonstrates the workflow for implementing a roadside traffic monitoring system using a single millimeter-wave radar: first, raw millimeter-wave radar signals are acquired; then, through point cloud extraction and enhancement processing, the radar signals are converted into dense 3D point cloud data. This enhanced point cloud serves as a unified input, supporting the following three core tasks:

[0124] (1) Traffic flow statistics based on continuous main beam crossing detection; (2) Vehicle speed estimation based on accurate passing anchor point detection; (3) High-resolution three-dimensional vehicle imaging based on conditional modified flow model under the guidance of enhanced point cloud representation and structure optimization mechanism.

[0125] A roadside traffic monitoring method based on a single millimeter-wave radar includes the following steps:

[0126] S1. Roadside vehicle data collection;

[0127] First, the millimeter-wave radar is fixedly installed on the side of the road, with the main beam direction basically perpendicular to the direction of vehicle travel, covering a multi-lane road area. The radar periodically transmits frequency-modulated continuous wave signals and receives echoes reflected from the vehicle surface and structure, forming multiple frames of raw intermediate frequency signal data.

[0128] S2. Millimeter-wave point cloud extraction and enhancement;

[0129] To address the aforementioned technical challenge of balancing point cloud visibility and temporal resolution, this invention proposes a point cloud extraction and enhancement method at the millimeter-wave radar raw signal processing level. This method aims to improve point cloud density and structural stability while maintaining the temporal sampling granularity. Existing methods for increasing point cloud density mainly include two strategies: multi-frame overlay and lowering the CFAR detection threshold. However, multi-frame overlay increases the number of points through temporal aggregation, which essentially sacrifices temporal resolution and can weaken the ability to characterize transient motion in high-speed vehicle scenarios. Lowering the detection threshold, while increasing the number of points, introduces a large amount of environmental clutter and multipath interference, disrupting the integrity of the point cloud structure and affecting the accuracy of subsequent volumetric statistics and velocity estimation. To solve these problems, this invention makes three improvements to the standard radar signal processing workflow.

[0130] (1) Chirp splitting-aggregation mechanism:

[0131] First, to improve the point cloud representation capability within a single frame, this invention constructs a chirp segmentation-aggregation mechanism. Let the single-frame radar signal be a two-dimensional matrix x(n,m), where... This represents the slow-time dimension chirp index. This represents the fast time-dimensional sampling point index. This invention performs segmentation processing on the original signal, proposing two candidate implementation methods: intra-chirp segmentation and inter-chirp segmentation, such as... Figure 2As shown. In the intra-chirp segmentation method, along the fast time dimension... Divide a single frame signal into Each sub-segment, mathematically expressed as:

[0132] ;

[0133] in, This segmentation method corresponds to coarse-scale sampling results in the distance dimension. In the inter-chirp segmentation method, along the slow time dimension... Divide a single frame signal into Each sub-segment, mathematically expressed as:

[0134] ;

[0135] in, This segmentation method corresponds to the coarse-scale sampling results in the velocity dimension. After segmentation, this invention constructs a signal set containing the original complete frame signal and the segmented sub-segment signals. ,in ,and Then, a Doppler-FFT operation is performed on each group of signals in the set to generate the corresponding distance-Doppler matrix:

[0136] ;

[0137] in:

[0138] ;

[0139] Through the above processing, this invention expands a single-resolution radar snapshot into multi-scale observation results within a single frame. Specifically, the complete frame signal provides fine-scale high-resolution results, intra-chirp segmentation provides coarse-scale observations in the range dimension, and inter-chirp segmentation provides coarse-scale observations in the velocity dimension. By fusing RDM results at different scales, multi-scale point cloud representation enhancement is achieved without reducing temporal resolution or introducing cross-frame overlay, thereby improving the detection capability of distant targets and weak echo targets in occluded scenes.

[0140] It should be noted that the computational complexity increases with the number of partitions. The growth is approximately linear. To ensure system real-time performance, this invention conducts a trade-off analysis on the segmentation method and parameters. Results show that inter-chirp segmentation outperforms intra-chirp segmentation in terms of temporal diversity; when the number of segmentations reaches a certain value... At this time, a better balance is achieved between computational complexity and point cloud enhancement effect. Therefore, the inter-chirp segmentation method is adopted in the preferred embodiment, and .

[0141] Figure 2 The diagram illustrates the intra-chirp and inter-chirp segmentation-aggregation mechanisms. Intra-chirp segmentation refers to dividing a single frame signal into several segments along the fast time dimension to obtain coarse-scale features in the distance dimension, while inter-chirp segmentation refers to dividing a single frame signal into several segments along the slow time dimension to obtain coarse-scale features in the velocity dimension.

[0142] (2) Adaptive soft threshold detection strategy:

[0143] Secondly, the target region of the vehicle is extracted from the generated RDM matrix. Addressing the issues of traditional CFAR's tendency to produce "self-masking" in strong echo regions and the introduction of noise through fixed low thresholds, this invention proposes an adaptive soft threshold detection method. Firstly, in the complete RDM... The core region is obtained by using a conservative threshold; then the signal-to-noise ratio (SNR), point volume (PV), and spatial compactness (SC) are calculated based on this region, and then fused using a fusion function. Generate joint quality evaluation indicators to adjust the adaptive coefficients:

[0144] .

[0145] in:

[0146] and These are used to avoid over-selection and under-selection, respectively.

[0147] SNR characterizes the prominence of the strongest target echo relative to background noise;

[0148] PV is the number of points in a single cluster after DBSCAN clustering, used to characterize the spatial scale of the target.

[0149] SC is the ratio of the number of points to the area of ​​its bounding box, used to measure the density of space.

[0150] Adopting a Sum-Product fusion strategy: ,in, This is the balance coefficient.

[0151] Ultimately based on Extract target regions from RDM at different scales. and These correspond to the mean and standard deviation of RDM at different scales.

[0152] (3) Bartlett beamforming extraction angle:

[0153] Then, the target distance-Doppler points extracted from RDMs of different scales in step (2) are angle-estimated to recover their three-dimensional spatial positions. To address the limitations of the angle resolution and sensitivity to multipath interference in traditional Angle-FFT methods, this invention employs a Bartlett beamforming algorithm based on a multi-antenna array for angle estimation. While existing high-resolution beamforming methods (such as Capon and MUSIC) can provide finer angle resolution, they require accumulating multiple chirps to construct a stable covariance matrix. This process is essentially time-domain aggregation, which reduces temporal resolution and is unsuitable for roadside real-time traffic monitoring scenarios.

[0154] Bartlett beamforming eliminates the need to construct a covariance matrix across chirps, enabling angle estimation within a single frame of data. This allows for stable angle measurement while maintaining consistent temporal sampling granularity. Specifically, for the target point set obtained in step (2), a direction vector is constructed using the received signal from the antenna array, and the azimuth angle of each point is obtained through Bartlett spatial spectrum search. Pitch angle ;

[0155] Calculate the three-dimensional coordinates by combining the corresponding distance information (range). :

[0156] ;

[0157] This allows us to recover the three-dimensional spatial distribution of the target points, forming a single-frame three-dimensional point cloud.

[0158] Finally, to obtain a vehicle-level point cloud set, this invention first spatially partitions the 3D point cloud based on lane geometry; then, within each lane region, the DBSCAN clustering algorithm is used to cluster points belonging to the same vehicle, resulting in an independent vehicle point cloud set. }, each of which This represents the 3D point cloud data of a vehicle.

[0159] S3. Traffic flow statistics and vehicle speed estimation based on continuous frame point clouds;

[0160] Traffic flow statistics:

[0161] After completing vehicle-level point cloud extraction and enhancement, this invention further utilizes the extracted vehicle point cloud for traffic flow statistics and addresses the "cross-lane occlusion" problem described in Part Two. Existing millimeter-wave radar traffic flow statistics methods often employ a bird's-eye view deployment, setting predefined detection lines within the radar's effective field of view. Counting is achieved by determining whether vehicles cross the detection lines, and errors are reduced by combining point cloud volume and trajectory length constraints. However, in roadside lateral observation scenarios, due to mutual occlusion between vehicles, vehicle point clouds may be incomplete or even temporarily missing. Even with the aforementioned point cloud enhancement methods, when large vehicles occlude smaller vehicles, effective vehicle centroids may still fail to be extracted for several consecutive frames, leading to counting interruptions or miscounting. Therefore, this invention proposes a roadside traffic flow statistics method with an occlusion compensation mechanism, such as... Figure 3 As shown.

[0162] First, the direction of the radar's main beam is set as the detection line, and the radar's effective field of view is defined as the detection area. For each frame... The point cloud of each vehicle will be extracted. Divided into Within each lane area, the centroid of the corresponding vehicle cluster is then calculated. And assign lane numbers to each centroid. .

[0163] Next, a Kalman filter is used to track vehicle trajectories in consecutive frames. The centroid of the current frame is matched with existing trajectories using the Hungarian algorithm; unmatched centroids initialize new trajectories based on their geometric positions and lane information, thereby ensuring continuous tracking of newly appearing vehicles.

[0164] For each trajectory, it is checked whether it intersects with the detection line. When an intersection is detected, the corresponding lane is recorded, and the traffic flow is incremented. By counting the number of vehicles passing through the detection line in each lane within a certain time period, the traffic flow for that time period is obtained.

[0165] However, trajectory interruptions can still occur when vehicles are occluded. When a vehicle is severely obstructed by a large vehicle in front, its point cloud may be extremely sparse over several consecutive frames. Even after point cloud enhancement, clustering algorithms may still be unable to extract effective vehicle centroids, leading to trajectory interruptions.

[0166] To compensate for the lack of such trajectories, this invention is based on historical trajectories. Using a Kalman filter to detect the obscured vehicle The missing centroids are continuously predicted to obtain the predicted centroids. When the new effective vehicle center of gravity When detected, the prediction process is stopped, and the predicted centroid and the new centroid are sequentially connected to the historical trajectory to recover the complete trajectory.

[0167] In complex and dynamic road environments, noise point clouds may cause the aforementioned occlusion compensation mechanism to fail in two situations.

[0168] The first scenario is that historical trajectories may originate from continuous noisy point clouds rather than actual vehicle point clouds. Direct prediction compensation in this case could lead to spurious trajectories and overcounting. To avoid this problem, this invention performs occlusion consistency checks on the compensated trajectories. For the predicted centroid... Detect whether there is a vehicle ahead. ,satisfy:

[0169] Furthermore, the spatial area occupied by the vehicle point cloud cluster With radar position The line connecting to the predicted centroid Existence of spatial intersection: .

[0170] in, This represents the set of the centroids of all vehicles. Indicates vehicle Located in the lane in front of the vehicle awaiting compensation; Indicates radar location The line connecting the predicted centroid; This represents the spatial region occupied by the point cloud cluster. The predicted centroid is determined to satisfy the physical occlusion relationship and is retained only if the above conditions are met; otherwise, compensation is terminated to avoid creating false trajectories.

[0171] The second scenario is: newly detected centroids. It may also originate from noisy point clouds. If mistakenly included in the trajectory, it could lead to trajectory deviation and incorrect counting. Therefore, this invention further compares the distance between the new centroid and the Kalman-predicted centroid:

[0172] ;

[0173] If this distance is smaller than the historical trajectory Average width of midpoint cloud If the observation is valid, the new centroid is considered a valid observation and is included in the trajectory; otherwise, the centroid is discarded.

[0174] Vehicle speed estimation

[0175] After enhancing point cloud extraction and vehicle trajectory continuity compensation, vehicle speed is further estimated based on the enhanced vehicle-level point cloud data. In roadside deployment scenarios, radars are typically installed perpendicular to the traffic flow direction, and the speed directly output by the radar is the radial speed relative to the radar position, rather than the vehicle's true longitudinal speed along the travel direction. Therefore, it is necessary to reconstruct the vehicle's true speed. Existing speed measurement methods based on millimeter-wave radar mostly rely on Doppler spectrum analysis or phase change to extract radial speed. However, these methods have inherent limitations: Doppler spectrum analysis suffers from speed ambiguity; and the azimuth information used for speed decomposition is easily affected by clutter in complex road environments, leading to large calculation errors and making it difficult to meet the high-precision speed measurement requirements of roadside systems.

[0176] To avoid speed ambiguity and noise interference, this invention employs a speed measurement principle based on the ratio of known spatial distance to transit time. It calculates the vehicle's transit time between two adaptively determined spatial anchor points and combines this with the distance between the anchor points to obtain the vehicle's true speed. However, in practical applications, determining the anchor point positions to accurately obtain the transit time is a core challenge. One direct method is to pre-set two fixed anchor point positions and extract frame differences when the centroid of the point cloud passes through the anchor points. However, due to the limited range-angle resolution of millimeter-wave radar and the influence of multipath effects, the centroid position fluctuates significantly during vehicle movement, leading to large speed measurement errors. Therefore, this invention does not use fixed anchor points but proposes a "coarse-to-fine" vehicle head alignment method to adaptively select anchor point positions. Figure 4 As shown, the intuitive basis of this method is that the vehicle head is the foremost part of the vehicle body, and its displacement can directly reflect the vehicle's actual longitudinal travel distance; in contrast, the position of the center of mass is easily affected by changes in the point cloud distribution.

[0177] Based on this, we propose a vehicle head alignment method to set anchor points. Specifically: First, two symmetrical coarse selection regions are defined on both sides of the radar to ensure sufficient clearance. Then, a fine selection is performed within each coarse selection region, and the vehicle head position with the highest point cloud quality is selected to determine the final anchor point. At this point, the problem becomes how to evaluate the quality of the vehicle head point cloud.

[0178] To address this issue, we utilize the following observation: vehicle fronts are generally smooth and geometrically regular. Therefore, we assign a quality score to the vehicle front point cloud to reflect the quality of each fine-grained candidate anchor point location. The specific steps are as follows:

[0179] First, the vehicle head is segmented by extracting points in the vehicle point cloud cluster whose X-axis coordinates are within 0.5m in front of the vehicle body.

[0180] Subsequently, the head quality score was defined:

[0181] ;

[0182] The rating consists of two parts:

[0183] 1) 2) Measure the smoothness of the head point cloud in the Z-axis direction (i.e., the degree of uniform distribution in the vertical direction). The regularity of the vehicle's front end in the X-axis direction (i.e., the tightness in the horizontal direction) is measured.

[0184] The calculation formula is as follows:

[0185] ;

[0186] in:

[0187] This represents the standard deviation of the head point in the Z-axis coordinate system.

[0188] This represents the total number of head points;

[0189] Indicates the first The X-axis coordinates of each point;

[0190] The X-axis coordinate represents the current refined candidate position.

[0191] In obtaining the fine location of each candidate Then, select from the left and right coarse selection areas respectively. The highest position is used as the final anchor point. It should be noted that in cases of occlusion between vehicles, the sparse point cloud will affect the candidate position. Similarly, the score is low. Therefore, this scoring mechanism can naturally take into account head alignment issues under occlusion conditions.

[0192] Finally, as the vehicle passes through two anchor points in sequence, the corresponding frame difference is calculated and converted into a time interval. The vehicle speed is obtained by dividing the distance between the anchor points by this time interval.

[0193] Figure 3 This demonstrates a roadside environment where a large vehicle (V1) ahead occludes a small vehicle (V2) behind, resulting in a sparse point cloud and even making it impossible to extract valid vehicle centroids. To complete the trajectory, the system uses Kalman filtering to predict the missing centroids (e.g., predicting...). and ), and when a new centroid appears (e.g. The predicted values ​​are sequentially connected with the actual centroids to form a complete trajectory, while occlusion checks are performed to determine the predicted centroids. Is it related to the point cloud region of the preceding vehicle V1? Intersection, that is, the line connecting the radar to the center of mass. Intersection with the point cloud region of the preceding vehicle) and reliability verification of the new centroid (comparing the distance between the new centroid and the predicted centroid). Average point cloud width of historical trajectories ;like but (If deemed valid, otherwise ignored) Ensure vehicle trajectories are continuous and counts are accurate.

[0194] Figure 4 This paper demonstrates a coarse-to-fine front-end alignment method used in vehicle speed estimation. First, symmetrical coarse regions are defined on both sides of the radar to ensure sufficient clearance. Then, fine-grained front-end positions are selected within each coarse region, and the optimal position is chosen as the anchor point for speed calculation by calculating the quality score of the front-end point cloud. Finally, the vehicle's true speed is obtained by calculating the distance d between two anchor points and the corresponding time interval t2-t1.

[0195] S4. Vehicle 3D Imaging Based on Enhanced Point Cloud.

[0196] The model training steps are as follows:

[0197] Due to bandwidth and antenna array limitations of millimeter-wave radar, the enhanced point cloud still suffers from insufficient point count. Furthermore, specular reflection and occlusion between vehicles further reduce the number of effective points.

[0198] 1) Conditional augmentation representation methods for sparsity problems:

[0199] To construct a unified and information-complete conditional input, this invention projects the original point cloud into a unified two-dimensional representation. Single projection methods suffer from information loss: bird's-eye view projection retains the top-down structure but compresses height information; distance map projection retains forward and side-view information but has weak representation capabilities for distant targets. Therefore, this invention employs a method of stitching and fusing bird's-eye view projection and distance map projection to construct an enhanced conditional representation:

[0200] From Gaussian noise distribution To the actual vehicle point cloud distribution Linear generation path:

[0201] ;

[0202] in, Indicates the sampling time step;

[0203] Linear path by velocity field Drive, to intermediate state The flow is directed towards the target distribution; a U-Net structure is used to predict the velocity field, and a regression approximation is performed on the true velocity field.

[0204] ;

[0205] An enhanced conditional representation is constructed by stitching together bird's-eye view projection and distance map projection:

[0206] ;

[0207] The network optimization objective is to minimize the difference between the predicted velocity field and the true velocity field. distance:

[0208] ;

[0209] in, This represents conditional guidance information constructed from radar point clouds;

[0210] Subsequently, the enhanced conditional input to the U-Net is embedded in multiple layers. To avoid the attenuation of conditional information during deep propagation, the conditional features are encoded as multiple independent inputs. This is injected into each layer of U-Net to achieve multi-scale structural constraints; simultaneously, the time steps are... By introducing conditional representations, time-aware embedding is achieved through scale modulation and offset modulation mechanisms:

[0211] This enhances the model's ability to perceive the location of the generated path;

[0212] Indicates by parameters The velocity field function is characterized and used in the context of conditional information. State variables are given under constraints In time The instantaneous rate of change of the point cloud along the generation path is output as a vector of changes in the state variables, used to determine the point cloud state at the time step. The direction and magnitude of updates within the system.

[0213] (ii) Imaging structure optimization methods for structural disorder problems:

[0214] The original conditional correction flow relies solely on velocity regression for optimization. Small deviations in the velocity field can accumulate and amplify during multi-step ODE iterations, leading to structural distortion. To improve imaging structural consistency, this invention introduces a reconstruction regression loss term in addition to the velocity regression loss:

[0215] ;

[0216] Through derivation, we can obtain:

[0217] ;

[0218] : Represents the reconstruction loss weighting coefficient, which is a positive number used to adjust the reconstruction loss. The degree of influence in the overall loss function is used to control the proportion of the output point cloud reconstruction loss that contributes to the model optimization process;

[0219] : Represents the point cloud state at the target time predicted by the model, i.e., the state at the current time. In the velocity field The reconstructed point cloud results obtained under the influence of the function are used to compare with the real point cloud. Calculate the reconstruction error.

[0220] The aforementioned weighting function exhibits an inverted U-shaped distribution, assigning higher weights at intermediate time steps. This helps the model learn a straighter generation path, improving structural fidelity. Furthermore, this invention introduces a learnable structure gating module into the conditional embedding, focusing on enhancing the main vehicle region. This concentrates the imaging process on the vehicle structure itself, suppressing background interference and improving the accuracy of detail reconstruction.

[0221] (iii) High-confidence point guided reasoning methods for noise problems:

[0222] Millimeter-wave radar point clouds are subject to multipath effects and contain noisy points. If the inference is guided by noise in the early stages, the error will amplify with each iteration. To address this, this invention proposes a high-confidence-point guided inference strategy.

[0223] During the inference phase, high confidence points with intensity values ​​in the top 15% are selected to construct the reference matrix. and with initial noise Fusion:

[0224] ;

[0225] in, It is a binary mask matrix. The matrix is ​​composed of all 1s, and the velocity field is calibrated simultaneously.

[0226] ;

[0227] This is used to control the guiding strength of high confidence points in the early stages of inference, and to gradually weaken them in the later stages;

[0228] From initial noise Set off, according to the distance from the walk Gradually updated:

[0229] ;

[0230] Indicates by parameters The velocity field function is characterized and used in the context of conditional information. State variables are given under constraints In time The instantaneous rate of change of the point cloud along the generation path is output as a vector of changes in the state variables, used to determine the point cloud state at the time step. The direction and magnitude of updates within the system.

[0231] The reconstructed three-dimensional structure of the vehicle is obtained through a finite number of iterations.

[0232] Finally, the generated results are back-projected into a high-fidelity 3D vehicle point cloud, achieving detailed 3D vehicle imaging.

[0233] Figure 5 Demonstrates the use of radar-extracted point clouds as conditional guidance The vehicle imaging process. The model learns from Gaussian noise during the training phase. To real vehicle point cloud The model employs a deterministic mapping and progressively reconstructs the vehicle's 3D geometry during the inference phase. To address the issues of sparse, disordered, and noisy point clouds, the model utilizes multi-layer conditional embedding to enhance point cloud information, structural optimization constraints to improve internal arrangement, and introduces high confidence points (HCP) guidance during inference to ensure that the generated 3D vehicle point cloud is highly consistent with the real vehicle in terms of structure and detail.

[0234] To verify the effectiveness of this patented method in comprehensive traffic monitoring tasks, representative existing methods were selected as comparative solutions for three functions: traffic flow statistics, vehicle speed estimation, and vehicle imaging, as detailed below:

[0235] Traffic flow statistics comparison method:

[0236] In traffic flow statistics tasks, two typical methods are selected for comparison. The first type is the traditional method based on range-Doppler maps (RDM) for direct target detection and counting. This method extracts targets and counts vehicles within the RDM space, representing a classic radar statistics scheme. The second type is a traffic flow statistics method based on a millimeter-wave radar top-view (BEV perspective). This method converts radar data to a top-view plane for vehicle detection and counting, but does not model the occlusion effect between vehicles. These two types of methods respectively represent the traditional statistical scheme and the improved scheme based on spatial mapping.

[0237] Vehicle speed estimation comparison method:

[0238] In the vehicle speed estimation task, two common speed estimation mechanisms are selected for comparison. The first is a speed estimation method based on Doppler Fast Fourier Transform (Doppler-FFT), which extracts the radial velocity information of the target through frequency domain analysis. The second is a speed estimation method based on phase change, which calculates the radial velocity of the target by analyzing the phase difference between adjacent samples, and then processes the radial velocity with the angle to obtain the true velocity. These two methods represent the mainstream speed measurement technology paths based on spectrum analysis and phase information, respectively.

[0239] Vehicle imaging comparison methods:

[0240] In traffic flow imaging tasks, to comprehensively evaluate the quality of 3D reconstruction, various methods were selected for comparison. These include: traditional radar target detection methods based on the Constant False Alarm Rate (CFAR) algorithm; deep generative model methods, such as Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs); the neural network model RPDNet specifically designed for radar imaging tasks; and the diffusion-based imaging method RangeDiff. These methods cover traditional signal processing methods, deep generative model methods, and diffusion-based generative model methods, enabling comparative analysis of 3D imaging performance from different technical perspectives.

[0241] Evaluation indicators:

[0242] To comprehensively evaluate the system performance of this invention, three independent quantitative indicators are used to assess vehicle traffic statistics, vehicle speed estimation, and high-precision imaging capabilities.

[0243] 1. Vehicle traffic flow statistics performance:

[0244] To evaluate the effectiveness of vehicle traffic flow statistics, this invention uses the accuracy rate (Acc) of vehicle traffic flow counting. v As an evaluation indicator, it is defined as follows:

[0245] ;

[0246] in, To estimate the number of vehicles, This is the actual number of vehicles counted. This normalized metric penalizes both overcounting and undercounting. Evaluation results show that Acc... v The higher the value, the better the performance of the vehicle traffic flow statistics.

[0247] 2. Vehicle speed estimation performance:

[0248] To evaluate vehicle speed estimation capability, this invention uses the average estimation error (MAE) as an evaluation index, which is defined as follows:

[0249] ;

[0250] in, Let be the estimated speed of the i-th vehicle. The actual speed is represented by the MAE value. Evaluation results show that the smaller the MAE value, the better the vehicle speed estimation performance.

[0251] 3. Vehicle imaging performance:

[0252] To evaluate radar imaging accuracy, this invention reconstructs the point cloud. Ground-based point cloud measured by lidar The following four types of indicators were used for comparison:

[0253] ChamferDistance (CD): Used to evaluate the degree of matching of local geometric details;

[0254] EarthMover's Distance (EMD): Used to assess the degree of global structural alignment;

[0255] Modified Hausdorff Distance (MHD): Used to detect spatial outliers and worst-fit bias;

[0256] F-Score: Used to jointly evaluate imaging accuracy and recall.

[0257] Evaluation results show that the method of the present invention has lower values ​​in CD, EMD, and MHD indicators, and higher F-Score, indicating that the radar imaging reconstruction accuracy is higher and can achieve high-fidelity vehicle point cloud reconstruction.

[0258] Traffic flow statistics results:

[0259] Comparison with existing methods:

[0260] Table 1: Accuracy Comparison Results with Existing Methods for Traffic Flow Statistics

[0261] .

[0262] As shown in Table 1, the accuracy of this patented method in traffic flow statistics is 97.8%, which is higher than that of the RDM method (36.2%) and the BEV method (93.2%).

[0263] The impact of vehicle obstruction on traffic flow statistics:

[0264] In multi-lane scenarios, vehicles in the inner lanes (leftmost lanes 1-2) are more susceptible to obstruction by vehicles in the outer lanes. Therefore, the accuracy of traffic flow counting for the inner and outer lanes was statistically analyzed separately. Figure 6 As shown:

[0265] The accuracy rate for counting inner lanes was 97.0%; the accuracy rate for counting outer lanes was 99.5%.

[0266] Test results show that although vehicles in the inner lane are more susceptible to obstruction, their statistical accuracy remains above 97%, only slightly lower than that of the outer lane.

[0267] The above results demonstrate that this method, through occlusion perception and compensation mechanisms, can effectively reduce statistical errors caused by mutual occlusion between vehicles and maintain high counting accuracy even in complex multi-lane environments.

[0268] The impact of different environmental conditions on traffic flow statistics:

[0269] Figure 7 As shown, traffic flow statistics tests were conducted in daytime, nighttime, rainy, and foggy conditions, with counting accuracy rates of 97.2%, 97.3%, 97.8%, and 98.1%, respectively. This indicates that the patented method can maintain stable performance under different lighting and weather conditions and has all-weather operation capability.

[0270] 2. Vehicle speed estimation results

[0271] Comparison with existing methods:

[0272] Table 2: Comparison of average estimation error with existing vehicle speed estimation methods

[0273] .

[0274] As shown in Table 2, the average vehicle speed estimation error of the method in this patent is only 3.40 km / h, which is significantly lower than that of the Doppler method (7.55 km / h) and the phase method (13.08 km / h), indicating that the method in this patent has a significant advantage in vehicle speed estimation accuracy.

[0275] The impact of vehicle occlusion on vehicle speed estimation:

[0276] like Figure 8 As shown, the average speed measurement errors of vehicles in the inner and outer lanes were statistically analyzed. The results indicate that even with some obstruction from vehicles in the inner lane, the speed measurement error remains within 5 km / h. The anchor point speed measurement mechanism employed in this patented method maintains stable estimation accuracy under partial signal obstruction, demonstrating excellent anti-obstruction capabilities.

[0277] Vehicle speed estimation under different environmental conditions:

[0278] like Figure 9As shown, speed measurement experiments were conducted under daytime, nighttime, rainy, and foggy conditions. The analysis results show that the patented method can maintain stable speed measurement performance under different lighting and weather conditions, with small error fluctuations, fully demonstrating the all-weather advantages of millimeter-wave radar systems.

[0279] Vehicle speed estimation at different speed levels:

[0280] like Figure 10 As shown, vehicles are categorized for testing according to speed ranges, including low speed (<60km / h), medium speed (60–80km / h), and high speed (>80km / h). Analysis results show that the speed measurement error increases moderately with increasing speed level, but remains below 5km / h at high speeds. The coarse-to-fine head alignment mechanism employed in this patented method effectively solves the problem of inaccurate speed measurement of high-speed targets, maintaining stable accuracy across different speed ranges.

[0281] 3. Vehicle imaging:

[0282] Comparison with existing methods:

[0283] Table 3: Quantitative comparison results of vehicle imaging with existing methods

[0284] .

[0285] The vehicle imaging module described in this application achieves high accuracy in vehicle 3D imaging tasks, with an average Chamfer Distance (CD) of 0.21, an Earth Mover's Distance (EMD) of 0.41, a Modified Hausdorff Error (MHE) of 0.22, and an F-Score of 74.1%.

[0286] The proposed imaging model was compared with typical generative models (including CFAR, cVAE, cGAN, RPDNet, and diffusion models). As shown in Table 3, this patent outperforms the comparison methods in all evaluation metrics. In particular, compared to the best-performing comparison method (RangeDiff), this patent reduces the average score by 0.67 across the three distance metrics and improves the F-Score by more than 40 percentage points. Figure 11 As shown in the visualization results, the present invention can achieve the most refined vehicle-level point cloud imaging effect, and can more accurately reconstruct the vehicle shape and geometry compared with existing methods. The above results fully demonstrate that the present invention, through the correction flow framework combined with the optimization module, can generate high-fidelity 3D vehicle images from sparse radar point clouds, exhibiting significant technical advantages.

[0287] In this specification, the present invention has been described with reference to specific embodiments. These embodiments are preferred embodiments of the present patent and are not intended to limit the scope of the invention. It should be noted that the present invention is not limited to the specific embodiments described above. Improvements, variations, combinations, substitutions, etc., made by those skilled in the art without departing from the principles of the present invention are all within the scope of protection claimed in the claims of the present invention.

[0288] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A roadside traffic monitoring method based on a single millimeter-wave radar, characterized in that, Includes the following steps: S1. Roadside vehicle data collection; S2. Millimeter-wave point cloud extraction and enhancement; S3. Traffic flow statistics and vehicle speed estimation based on continuous frame point clouds; The steps for traffic flow statistics are as follows: The radar main beam direction is set as the detection line, and the radar effective field of view is defined as the detection area: for each frame The point cloud of each vehicle will be extracted. Divided into In each lane area; Calculate the centroid of the corresponding vehicle cluster And assign lane numbers to each centroid. ; A Kalman filter is used to track vehicle trajectories in consecutive frames. The centroid of the current frame is matched with existing trajectories using the Hungarian algorithm. Unmatched centroids initialize new trajectories based on their geometric positions and lane information, thereby ensuring continuous tracking of newly appearing vehicles. For each trajectory, check whether it intersects with the detection line: when the trajectory intersects with the detection line, record the corresponding lane and increment the traffic flow by 1. By counting the number of vehicles passing through the detection line in each lane within a certain time period, obtain the traffic flow for that time period. The steps for estimating vehicle speed are as follows: The vehicle's front end is segmented by extracting points from the vehicle's point cloud cluster whose X-axis coordinates lie within a predetermined range in front of the vehicle body; a head quality score is defined. ； Measure the smoothness of the head point cloud in the Z-axis direction; The regularity of the vehicle's front end along the X-axis is measured; the formula for its calculation is: ； in: This represents the standard deviation of the head point in the Z-axis coordinate system. This represents the total number of head points; Indicates the first The X-axis coordinates of each point; The X-axis coordinates of the current refined candidate position; In obtaining the fine location of each candidate Then, select from the left and right coarse selection areas respectively. The highest position is used as the final anchor point; As the vehicle passes through two anchor points in sequence, the corresponding frame difference is calculated and converted into a time interval. The vehicle speed is obtained by dividing the distance between the anchor points by this time interval. S4. Vehicle 3D Imaging Based on Enhanced Point Cloud; The input point cloud is converted into a bird's-eye view and a distance map representation, which are then stitched together to form a unified conditional feature. Subsequently, Gaussian noise is used as the initial generation state, and the initial noise distribution is partially replaced and constrained by high-confidence millimeter-wave points with strong reflection to construct a calibrated initial state to enhance consistency with the real vehicle structure. The conditional feature is then input into the imaging network to achieve 3D vehicle imaging output.

2. The roadside traffic monitoring method based on a single millimeter-wave radar according to claim 1, characterized in that, Step S2: Millimeter-wave point cloud extraction and enhancement. The specific steps are as follows: S21. Reconstruct the single-frame data into a slow-time–fast-time two-dimensional matrix, segment the original frame data according to the fast-time dimension to obtain multiple sub-segments; perform distance and Doppler processing on the complete frame and each sub-segment to form a multi-scale distance-velocity map. S22. Calculate an adaptive soft threshold based on the complete frame results, perform target filtering on the results at each scale, and extract the effective target points on the distance-velocity map; S23. Use multi-antenna Bartlett beamforming to estimate the azimuth and elevation angles of the target point, and combine them with range information to convert them into three-dimensional spatial coordinates; S24. Perform lane partitioning and density clustering on the generated 3D point cloud to obtain the enhanced point cloud corresponding to each vehicle.

3. The roadside traffic monitoring method based on a single millimeter-wave radar according to claim 2, characterized in that, In step S21, the original signal is segmented using intra-chirp segmentation and / or inter-chirp segmentation. Intra-chirp segmentation: Let a single-frame radar signal be a two-dimensional matrix x(n,m), where This represents the slow-time dimension chirp index. Represents the index of the sampling point in the fast time dimension, along the fast time dimension. Divide a single frame signal into Each sub-segment, mathematically expressed as: ； in, This segmentation method corresponds to the coarse-scale sampling results of the distance dimension; In the inter-chirp partitioning approach, along the slow time dimension Divide a single frame signal into Each sub-segment, mathematically expressed as: ； in, This segmentation method corresponds to the coarse-scale sampling results of the velocity dimension; After segmentation, the signal set includes the original complete frame signal and the segmented sub-segment signals. ,in ,and Then, a Doppler-FFT operation is performed on each group of signals in the set to generate the corresponding range-Doppler RDM matrix: ； in: 。 4. The roadside traffic monitoring method based on a single millimeter-wave radar according to claim 3, characterized in that, Step S22 involves extracting the target region of the vehicle from the generated RDM matrix, specifically as follows: First, in the original, unsegmented RDM graph The core region is obtained by using a conservative threshold. Subsequently, through the fusion function Generate joint quality evaluation indicators to adjust the adaptive coefficients: ； in: and These are used to avoid over-selection and under-selection, respectively. SNR characterizes the prominence of the strongest target echo relative to background noise; PV is the number of points in a single cluster after DBSCAN clustering, used to characterize the spatial scale of the target. SC is the ratio of the number of points to the area of ​​its bounding box, used to measure the density of space. Adopting a Sum-Product fusion strategy: ； in, This is the balance coefficient; Ultimately based on Extract target regions from RDM at different scales. and These correspond to the mean and standard deviation of RDM at different scales.

5. The roadside traffic monitoring method based on a single millimeter-wave radar according to claim 1, characterized in that, In step S3, under the condition of occlusion between vehicles, based on historical trajectories... Using a Kalman filter to detect the obscured vehicle The missing centroids are continuously predicted to obtain the predicted centroids. When the new effective vehicle center of gravity When detected, the prediction process is stopped, and the predicted centroid and the new centroid are sequentially connected to the historical trajectory to recover the complete trajectory.

6. The roadside traffic monitoring method based on a single millimeter-wave radar according to claim 1, characterized in that, In step S3, noise point cloud processing: The first scenario is that the historical trajectory may originate from continuous noisy point clouds, rather than actual vehicle point clouds, which affects the prediction of the centroid. Detect whether there is a vehicle ahead. ,satisfy: Furthermore, the spatial area occupied by the vehicle point cloud cluster With radar position The line connecting to the predicted centroid Existence of spatial intersection: ; in, This represents the set of the centroids of all vehicles. Indicates vehicle Located in the lane in front of the vehicle awaiting compensation; Indicates radar location The line connecting the predicted centroid; The predicted centroid is determined to satisfy the physical occupancy relationship and is retained only when the above conditions are met; otherwise, compensation is terminated to avoid the formation of false trajectories. The second scenario is: newly detected centroids. Also derived from noisy point clouds, compare the distance between the new centroid and the Kalman-predicted centroid: ； If this distance is smaller than the historical trajectory Average width of midpoint cloud If the observation is valid, the new centroid is considered a valid observation and is included in the trajectory; otherwise, the centroid is discarded.

7. The roadside traffic monitoring method based on a single millimeter-wave radar according to claim 1, characterized in that, Step S4 involves performing 3D vehicle imaging based on the enhanced point cloud, including model training and model inference: The model training steps are as follows: From Gaussian noise distribution To the actual vehicle point cloud distribution Linear generation path: ； in, Indicates the sampling time step; Linear path by velocity field Drive, to intermediate state The flow is directed towards the target distribution; a U-Net structure is used to predict the velocity field, and a regression approximation is performed on the true velocity field. ; An enhanced conditional representation is constructed by stitching together bird's-eye view projection and distance map projection: ; The network optimization objective is to minimize the difference between the predicted velocity field and the true velocity field. distance: ; in, This represents conditional guidance information constructed from radar point clouds; The conditional input to the U-Net is embedded in multiple layers. To avoid the attenuation of conditional information during deep propagation, the conditional features are encoded as multiple independent inputs. This is injected into each layer of U-Net to achieve multi-scale structural constraints; simultaneously, the time steps are... Introducing conditional representations, time-aware embedding is achieved through scale modulation and offset modulation mechanisms: ； and They represent the parameters respectively. The scaling and offset functions are represented and used to scale the time variable. Conditional features Perform linear modulation; During the inference phase, a reference matrix is ​​constructed by selecting strength values ​​located at high confidence points. and with initial noise Fusion: ； in, It is a binary mask matrix. The matrix is ​​composed of all 1s, and the velocity field is calibrated simultaneously. ； It is used to control the guiding strength of high confidence points in the early stages of inference; From initial noise Set off, according to the distance from the walk Gradually updated: ； Indicates by parameters The velocity field function is characterized and used in the context of conditional information. State variables are given under constraints In time The instantaneous rate of change of the point cloud along the generation path is output as a vector of changes in the state variables, used to determine the point cloud state at the time step. The direction and magnitude of updates within; Through iteration, the reconstructed three-dimensional structure of the vehicle is obtained.

8. The roadside traffic monitoring method based on a single millimeter-wave radar according to claim 7, characterized in that, Based on the velocity regression loss, a reconstruction regression loss term is introduced: ； Through derivation, we can obtain: ； : Represents the reconstruction loss weighting coefficient, which is a positive number used to adjust the reconstruction loss. The degree of influence in the overall loss function is used to control the proportion of the output point cloud reconstruction loss contributing to the model optimization process; : Represents the point cloud state at the target time predicted by the model, i.e., the state at the current time. In the velocity field The reconstructed point cloud results obtained under the influence of the function are used to compare with the real point cloud. Calculate the reconstruction error.

9. A roadside traffic monitoring system based on a single millimeter-wave radar, characterized in that, It includes 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 3D imaging module; Data acquisition module: The millimeter-wave radar is fixedly installed on the side of the road, with the main beam of the radar perpendicular to the direction of vehicle travel, covering multi-lane road areas; The radar periodically transmits frequency-modulated continuous wave signals and receives the echoes reflected from the vehicle surface and structure, forming multiple frames of raw intermediate frequency signal data. The millimeter-wave point cloud extraction and enhancement module first reconstructs single-frame data into a slow-time–fast-time two-dimensional matrix. The original frame data is then segmented according to the fast-time dimension to obtain multiple sub-segments. Next, range and Doppler processing is performed on the complete frame and each sub-segment to generate a multi-scale range-velocity map. Then, an adaptive soft threshold is calculated based on the complete frame results, and target filtering is performed on the results at each scale to extract valid target points on the range-velocity map. Based on this, the azimuth and elevation angles of the target points are estimated using a multi-antenna Bartlett beamforming system, and combined with range information, converted into three-dimensional spatial coordinates. Finally, lane partitioning and density clustering are performed on the generated three-dimensional point cloud to obtain the enhanced point cloud corresponding to each vehicle. Traffic flow statistics module: converts the captured continuous frame raw signals of vehicles into continuous frame millimeter wave point cloud data; Based on the continuous frame point cloud, a detection area and detection line are set within the radar field of view. Lane division and clustering are performed on each frame point cloud, and the centroid positions of each vehicle point cloud cluster are extracted and calculated, while lane numbers are assigned. Subsequently, a vehicle state model is established based on Kalman filtering to predict and update the centroids of continuous frames. The Hungarian algorithm is used to match the current frame centroid with historical trajectories. Unmatched targets are initialized as new trajectories to achieve continuous vehicle tracking. After the trajectory is formed, it is determined whether each vehicle trajectory crosses the preset detection line. If a crossing occurs, the count is accumulated according to the lane to which it belongs. When occlusion causes a temporary loss of centroids, Kalman filtering is used for trajectory prediction and compensation. In subsequent frames, the re-emerging centroids are matched and updated for validity. Finally, the number of vehicles passing through the detection line within the set time window is counted to obtain the traffic flow results for the corresponding time period. Vehicle speed estimation module: Based on the enhanced continuous frame vehicle point cloud data, the vehicle speed is calculated through spatial displacement and time difference. First, two coarse anchoring regions are determined on both sides of the road. Then, the vehicle front point cloud is extracted as a candidate anchor point in each coarse region and matched with multiple preset fine positions in the coarse region to calculate a quality score. The score is determined by the dispersion of the vehicle front point in the vertical direction and the tightness of the fit along the driving direction. The fine position with the highest score is selected as the final spatial anchor point. Then, the time difference when the front of the same vehicle passes through two anchor points in sequence is recorded, and combined with the known spatial distance between the two anchor points, the average vehicle speed is obtained according to the ratio of distance to time. The 3D imaging module converts the input point cloud into a bird's-eye view and a distance map representation, respectively, and then stitches and fuses them to form a unified conditional feature. Subsequently, Gaussian noise is used as the initial generation state, and strong reflective high-confidence millimeter-wave points are used to partially replace and constrain the initial noise distribution to construct a calibrated initial state, thereby enhancing the consistency with the real vehicle structure. The conditional feature is input into the imaging network, and combined with time-step parameter modulation, the point cloud state is updated iteratively through discrete time steps, gradually approximating the real vehicle 3D structure under conditional guidance. Finally, the generated result is back-projected into 3D space to output the vehicle 3D imaging point cloud result.

Images