A congestion vehicle flow lightning vision fusion detection method and system with multi-feature synergistic enhancement

Abstract

The application discloses a congestion vehicle flow radar and vision fusion detection method and system with multi-feature synergistic enhancement, relates to the field of intelligent traffic perception, adopts a coaxial arrangement mode of a millimeter wave radar and a camera, constructs a layered space-time registration architecture, realizes high-precision space-time alignment of multi-source data, removes static clutter and extracts vehicle scattering characteristics in combination with a time-frequency analysis and a hybrid clutter suppression algorithm at a radar end, carries a multi-stage image preprocessing strategy at a visual end, combines a lightweight SSD model embedded with an attention mechanism, and improves target recognition capability. A vehicle flow density adaptive DBSCAN clustering algorithm is designed, and a speed and contour double constraint is used to optimize clustering segmentation effect; stable tracking is realized by relying on a CTRV dynamics model and a multi-feature cost correlation mechanism; a smooth and congested dual-mode vehicle flow counting logic is constructed, an integral correction mechanism is introduced to correct statistical errors of queue overflow; and finally, real targets are screened through dynamic weight fusion and multi-stage abnormality checking, and detection data is purified.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation perception technology, and in particular relates to a multi-feature collaborative enhancement method and system for radar-visual fusion detection of congested traffic flow. Background Technology

[0002] In urban traffic networks, key passageways such as intersections and highway ramps experience high traffic saturation, with frequent vehicle starts and stops, easily leading to persistent congestion. These conditions typically involve complex traffic characteristics such as large vehicles obstructing traffic, small vehicle spacing, and low-speed vehicle congestion, placing high demands on the all-weather detection capabilities, dense target segmentation capabilities, and environmental interference resistance of sensing equipment. Currently, mainstream traffic flow sensing methods include visual detection, millimeter-wave radar detection, and geomagnetic detection.

[0003] To overcome the limitations of single-sensor detection, radar-visual fusion has become the mainstream improvement solution in the current traffic perception field. However, existing fusion technologies still have significant technical shortcomings: First, the fusion level is shallow, mostly only completing a simple mapping between radar point clouds and image pixels, without achieving deep coupling at the feature level. The correlation between multi-source features is weak, and target mismatch problems are prone to occur in densely congested scenarios. Second, the feature mining dimension is singular, relying only on visual contour features and failing to fully utilize radar scattering features and vehicle dynamic motion features, resulting in insufficient vehicle representation capabilities under occlusion conditions. Third, low-speed adaptability is poor, and traditional radar clutter suppression algorithms easily filter out low-speed, stationary queuing vehicles, causing missed detections in congested traffic flows. Fourth, the clustering algorithm has insufficient adaptive capability, and conventional algorithm parameters are fixed, making it unable to adapt to the segmentation requirements of extremely small vehicle spacing under congested conditions. Fifth, the tracking stability is weak, and defects such as ID jumps and trajectory breaks are prone to occur in scenarios with frequent vehicle starts and stops and instantaneous occlusion. Sixth, the counting logic is fixed, generally adopting a single cross-line counting mode, which cannot identify stagnant queuing vehicles, resulting in large statistical errors in congested traffic flows. In addition, most roadside detection radars on the market use the 77GHz frequency band, which has been restricted by the Ministry of Industry and Information Technology to be used exclusively for vehicles. Roadside deployment does not comply with radio management regulations, and there are compliance risks in the implementation of such projects.

[0004] Based on the above reasons, this invention designs a multi-feature collaborative enhancement method and system for radar-visual fusion detection of congested traffic flow. Summary of the Invention

[0005] The purpose of this invention is to solve the problems in the prior art, and to propose a multi-feature collaborative enhancement method and system for radar-visual fusion detection of congested traffic flow.

[0006] This invention first discloses a multi-feature collaborative enhancement method for radar-visual fusion detection of congested traffic flow, comprising the following steps: Acquire raw echo signals from millimeter-wave radar and image data captured by a camera; A hierarchical spatiotemporal synchronization architecture, which includes hardware time hard synchronization and dynamic spatial deviation correction, is adopted to synchronize and align the radar signal and the image in both time and spatial dimensions. The synchronized radar signal is subjected to time-frequency transformation, and a hybrid clutter suppression algorithm that includes dynamic adjustment of the suppression threshold based on traffic density is used to remove static environmental clutter and extract the radar scattering characteristics of vehicle targets. The synchronized image is processed using a lightweight target detection model with an embedded attention mechanism to extract the visual texture features of the vehicle target. The radar point cloud is projected onto the image coordinate system, and a clustering algorithm with adaptive adjustment of clustering neighborhood parameters according to traffic flow density is used to segment the fused features. The segmentation results are purified by speed consistency constraints and contour overlap constraints, and vehicle state parameters are fitted by weighted fusion. Motion state estimation is performed based on the vehicle kinematics model. Inter-frame target association matching is performed using a cost function that integrates position features, velocity features, and appearance features. Trajectory prediction compensation is set under occlusion conditions to achieve continuous tracking of multiple targets. The system determines the traffic congestion status based on the real-time vehicle speed, adaptively switches between the smooth cross-line counting mode and the congestion dwell counting mode, and performs supplementary vehicle counting for queue overflow scenarios. The confidence scores of radar detection and visual detection are adaptively weighted and fused to obtain a comprehensive confidence score. Valid targets are selected and false targets are eliminated based on the comprehensive confidence score threshold.

[0007] In the above method, the hardware time synchronization is: using the radar sampling clock as the global time reference, a hardware level-triggered synchronization mechanism is adopted to constrain the time error between the radar sampling time and the camera acquisition time within a preset allowable range.

[0008] In the above method, the dynamic spatial deviation correction is as follows: continuously extract stable feature points of the road at preset time intervals, and use the least squares method to iteratively optimize and update the spatial transformation matrix to compensate for the cumulative registration error caused by environmental disturbances.

[0009] In the above method, the hybrid clutter suppression algorithm includes: using inter-frame differential operation to remove static background clutter, using zero-frequency notch filtering to suppress ground DC low-frequency interference, and dynamically adjusting the constant false alarm threshold based on real-time traffic density.

[0010] In the above method, the lightweight object detection model with embedded attention mechanism is as follows: a channel attention module is embedded in the SSD object detection network, and the channel attention module calculates the attention weights of the feature channels through global average pooling and multilayer perceptron; the backbone network of the object detection model is simplified and compressed through convolutional layers.

[0011] In the above method, the adaptive adjustment of the clustering neighborhood parameters with traffic flow density specifically involves: characterizing traffic flow density by statistically analyzing the number of effective radar point clouds within a unit area; reducing the neighborhood radius of the DBSCAN clustering algorithm and increasing the minimum core point threshold under congested conditions, and expanding the neighborhood radius and decreasing the minimum core point threshold under smooth conditions; the speed consistency constraint ensures that the deviation between the movement speed of each point within a cluster and the average speed of the cluster does not exceed a preset threshold; and the contour overlap constraint ensures that the intersection-union ratio of the visual detection frames of each target within the same cluster is not lower than a preset threshold.

[0012] In the above method, the vehicle kinematic model is a constant turning rate and speed model; the cost function is composed of a weighted sum of position deviation cost, speed deviation cost and visual appearance similarity cost; the inter-frame target association matching adopts the Hungarian algorithm to solve the optimal matching relationship with the global minimum matching cost as the optimization objective.

[0013] In the above method, the trajectory prediction compensation under the occlusion state is as follows: when the detection signal of a single sensor is missing, the vehicle position is continuously predicted iteratively by relying on the redundant feature information of heterogeneous sensors combined with Kalman filtering; and, the number of consecutive effective feature accumulation frames is set as the condition for determining the new trajectory, and the number of consecutive disappearing frames is set as the condition for determining the trajectory extinction, so as to realize trajectory life cycle management.

[0014] In the above methods, the smooth crossing counting mode is as follows: traffic flow statistics are completed by using a trajectory crossing discrimination method. When a vehicle's trajectory crosses the boundary of the statistical area and the tracking identifier is not repeatedly recorded, it is determined as a single valid passage event. The congestion dwelling counting mode is as follows: queued vehicles are determined by the continuous dwelling time of vehicles in the statistical area, and the same tracking identifier only participates in the statistics once in a single congestion cycle. The queue overflow supplementary counting is as follows: a regional saturation capacity threshold is set. When the number of vehicles in the statistical area in real time is greater than the threshold, the scale of queued vehicles outside the field of view is extra-counted based on the upstream traffic flow entry rate to construct a fitting function.

[0015] Secondly, the present invention provides a multi-feature collaborative enhancement system for congested traffic flow radar-visual fusion detection, used in the above-mentioned method, comprising a coaxially mounted millimeter-wave radar and camera acquisition module, an embedded edge computing terminal, and a data storage and transmission module; the embedded edge computing terminal includes: The data acquisition and hierarchical spatiotemporal synchronization registration module is used to acquire radar signals and image data and perform synchronization and alignment using a hierarchical spatiotemporal synchronization architecture. The radar scattering feature extraction module is used to perform time-frequency transformation on radar signals and extract radar scattering features using a hybrid clutter suppression algorithm; The visual texture feature extraction module is used to extract visual texture features using a lightweight object detection model with an embedded attention mechanism. A multi-feature collaborative fusion clustering module is used for projection mapping, adaptive clustering, dual-constraint purification, and weighted fusion fitting of vehicle state parameters; A multi-feature joint multi-target tracking module is used for motion state estimation, multi-feature cost association matching, and occlusion compensation tracking; The congestion adaptive dual-mode traffic counting module is used for traffic flow status identification, adaptive switching of counting modes, and queue overflow supplementary counting. A multi-feature joint anomaly verification module is used for adaptive fusion of dual-source confidence and removal of false targets.

[0016] The beneficial effects of this invention are as follows: 1. This invention replaces software synchronization with hardware-level triggered hard synchronization, combined with rigid feature point spatial calibration and least squares method for dynamic updating of the homography matrix, to suppress long-term registration offset and feature mismatch in congested scenarios. It employs triple clutter suppression: inter-frame differencing, zero-frequency notch filtering, and adaptive constant false alarm threshold for traffic density. This preserves low-speed vehicle echoes amidst dense clutter, extracting four-dimensional features: distance, speed, azimuth, and scattering cross-section. Furthermore, it utilizes multi-level preprocessing combined with channel attention-based lightweight SSD to enhance occluded vehicle textures and simplify convolutional layers, balancing detection accuracy and edge deployment.

[0017] 2. This invention dynamically adjusts the DBSCAN neighborhood radius and the number of core points based on point cloud density. During congestion, the neighborhood radius is reduced to prevent clustering, while during smooth traffic, it is expanded to maintain clustering. Combined with dual constraints of velocity deviation and cross-union ratio (CUNR) to eliminate interfering points and suppress misclassification due to adhesion. A CTRV model is used to construct the state vector, fusing position, velocity, and appearance cost functions with the Hungarian algorithm for inter-frame matching to suppress ID jumps. An occlusion tolerance frame count is set, and continuous prediction is achieved using Kalman filtering, along with lifecycle management to eliminate false trajectories and ensure tracking continuity.

[0018] 3. This invention divides the sensing area into multiple regions, switches between cross-line counting and dwell counting based on speed thresholds, sets dwell time and deduplication constraints to prevent duplication, establishes a saturation capacity threshold, integrates and extra-far-end counting when the queue overflows, corrects statistical biases, and simultaneously constructs an adaptive weighted fusion model of radar and visual confidence, adjusts weights based on congestion status, sets confidence thresholds to filter real targets and eliminate false detections, and cleans abnormal data by combining trajectory constraints, outputting lane-level flow, vehicle speed, density, and vehicle type ratio. Attached Figure Description

[0019] Figure 1 This is a flowchart of a multi-feature collaborative enhancement method for traffic congestion detection using radar-visual fusion.

[0020] Figure 2 This is a hierarchical spatiotemporal synchronous registration architecture diagram of a multi-feature collaborative enhancement method for traffic congestion detection using radar-visual fusion.

[0021] Figure 3 This is a flowchart of radar scattering feature extraction in a multi-feature collaborative enhancement method for congested traffic flow radar-visual fusion detection disclosed in this invention.

[0022] Figure 4 This is a flowchart of visual texture feature extraction in a multi-feature collaborative enhancement method for traffic congestion detection using radar-visual fusion.

[0023] Figure 5 This is a module architecture diagram of a multi-feature collaborative enhancement traffic flow radar-visual fusion detection system disclosed in this invention. Detailed Implementation

[0024] To facilitate understanding of this application and to make the aforementioned objectives, features, and advantages of this application more apparent, a detailed description of specific embodiments of this application is provided below in conjunction with the accompanying drawings. Numerous specific details are set forth in the following description to provide a thorough understanding of this application, and preferred embodiments are shown in the accompanying drawings. However, this application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application. This application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified. In the description of this application, "several" means at least one, such as one, two, etc., unless otherwise explicitly specified. It should be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementations. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is only for describing particular implementations and is not intended to limit the scope of this application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0025] Reference Figures 1-5A multi-feature collaborative enhancement method for traffic congestion detection using radar-visual fusion includes the following steps: S1: First, acquire the raw echo signal from the millimeter-wave radar and the image data collected by the camera; A hierarchical spatiotemporal synchronization architecture, which includes hardware time hard synchronization and dynamic spatial deviation correction, is adopted to synchronize and align radar signals and images in both time and spatial dimensions.

[0026] The synchronized radar signal is subjected to time-frequency transformation, and a hybrid clutter suppression algorithm that includes dynamic adjustment of the suppression threshold based on traffic density is used to remove static environmental clutter and extract the radar scattering characteristics of vehicle targets.

[0027] A lightweight target detection model with an embedded attention mechanism is used to extract the visual texture features of vehicle targets from the synchronized images.

[0028] The radar point cloud is projected onto the image coordinate system, and a clustering algorithm with adaptive adjustment of clustering neighborhood parameters according to traffic flow density is used to segment the fused features. The segmentation results are then purified using speed consistency constraints and contour overlap constraints, and vehicle state parameters are fitted by weighted fusion.

[0029] Motion state estimation is performed based on the vehicle kinematics model. Inter-frame target association matching is performed using a cost function that integrates position features, velocity features, and appearance features. Trajectory prediction compensation is set under occlusion conditions to achieve continuous tracking of multiple targets.

[0030] The system determines the traffic congestion status based on the real-time vehicle speed, adaptively switches between the smooth cross-line counting mode and the congested stationary counting mode, and performs supplementary vehicle counting for queue overflow scenarios.

[0031] The confidence scores of radar detection and visual detection are adaptively weighted and fused to obtain a comprehensive confidence score. Valid targets are selected and false targets are eliminated based on the comprehensive confidence score threshold.

[0032] In one feasible embodiment, the millimeter-wave radar and high-definition camera are coaxially mounted to form an integrated acquisition device. The installation height of the device can be fixed at 5m, and the device depression angle can be set to 20° to ensure that the radar detection beam and the camera imaging field of view completely overlap, minimizing the detection blind zone. A hierarchical spatiotemporal synchronization registration architecture is adopted, which achieves high-precision alignment of radar and camera data through three-layer correction methods: hardware time synchronization, static spatial calibration, and dynamic deviation correction. This overcomes the shortcomings of traditional single software synchronization and fixed calibration matrix, such as large time delay, serious long-term offset, and low mapping accuracy.

[0033] (1) Hard synchronization of time: The radar reference clock is combined with the hardware level triggering synchronization mechanism, abandoning the traditional software triggering asynchronous acquisition method. The time error between the radar sampling time and the camera acquisition time is constrained within the preset allowable range, and the radar sampling time is set as the reference time. The camera capture time is Set the maximum allowable time error The hardware synchronization constraints satisfy:

[0034] Using the radar sampling clock as the global time reference, the sampling frame rate of the two sensors is unified at 10Hz. The hardware triggering method controls the time jitter within 3ms, eliminates the time misalignment problem caused by asynchronous acquisition, and ensures that the radar motion data and visual imaging data are strictly matched in the time dimension.

[0035] (2) Spatial registration: The checkerboard calibration method is used to solve the camera's internal distortion parameters. In this embodiment, the camera distortion coefficients are taken as: radial distortion coefficients. , Tangential distortion coefficient , The distortion correction model is as follows:

[0036] in: These are the coordinates of the distorted pixels; The extreme radius from the pixel to the image center is given. Based on this, rigid static feature points of the road, such as lane lines, curbs, and guardrails, which are not easily deformed, are selected as matching control points. Unlike traditional matching methods that rely on random natural features, this rigid feature point selection strategy can effectively avoid feature mismatch problems in scenarios with dense traffic flow and large vehicle occlusion. Accurate transformation from the radar coordinate system to the image coordinate system is achieved by solving the homography transformation matrix. In this embodiment, 25 sets of stable matching points are selected to solve the 3×3 homography transformation matrix. This completes the precise transformation from the radar coordinate system to the image coordinate system. The mapping relationship is as follows:

[0037] in: The coordinates are rectangular coordinates on the radar's ground plane. In this embodiment, the spatial mapping pixel error is controlled within 0.8 pixels, effectively avoiding feature mismatch problems caused by dense traffic flow and large vehicles obstructing the view.

[0038] (3) Dynamic Deviation Correction: In terms of dynamic deviation correction, stable road feature points are continuously extracted at preset time intervals. The homography transformation matrix is ​​updated iteratively using the least squares method to continuously compensate for the cumulative registration error caused by external environmental disturbances such as equipment vibration, temperature drift, and illumination changes. The iterative update time interval is set to 20s. Stable road feature points are periodically extracted, and the homography matrix is ​​updated iteratively using the least squares loss function. The expression of the loss function is:

[0039] in: For the coordinates of the corresponding feature points in the image, The coordinates of the feature points at the radar end; the number of feature points selected in a single iteration. The iterative convergence threshold is set to 0.0005. By minimizing the mapping error and dynamically correcting the transformation matrix, the registration offset caused by temperature drift and equipment vibration is continuously compensated, avoiding the decay of registration accuracy with system runtime and improving the long-term stability of system registration in complex outdoor traffic scenarios.

[0040] Through the coordinated use of the three-layer correction method of hardware time hard synchronization, spatial precise registration and dynamic deviation correction, high-precision spatiotemporal alignment of multi-source heterogeneous data is achieved, laying a precise foundation of homogeneous data for subsequent feature-level fusion.

[0041] In terms of radar data processing, millimeter-wave radar acquires real-time mid-frequency IQ raw complex sampling data of roads. A two-dimensional fast Fourier transform is used to convert the time-domain IQ sampling signal into a range-Doppler spectrum. Through joint analysis of the range and velocity dimensions, the echo amplitude distribution of targets at different distances and speeds can be intuitively depicted, laying a frequency domain data foundation for subsequent clutter identification and weak target extraction. In terms of clutter suppression, a triple suppression strategy is adopted, integrating inter-frame differential clutter cancellation, zero-frequency notch filtering, and adaptive threshold adjustment based on traffic flow. Inter-frame differential operations are used to remove static background clutter such as road surfaces and guardrails with constant amplitude. Combined with zero-frequency notch filtering, ground-based DC low-frequency interference is suppressed. Unlike traditional radar algorithms that use a single fixed threshold for constant false alarm rate (CFAR), this invention dynamically adjusts the CFAR threshold based on real-time traffic density. Under congested conditions, the threshold is appropriately increased to avoid false filtering of low-speed vehicle echoes, while under smooth conditions, a moderate threshold is maintained to effectively suppress clutter. This adaptive threshold adjustment method effectively solves the problem of low-speed vehicle echoes being easily overwhelmed by clutter in congested scenarios, significantly improving anti-interference capabilities in complex road environments. Motion parameters are calculated from the clean echo signal after clutter suppression. The radial velocity of the vehicle is determined based on the Doppler effect. Simultaneously, four types of scattering features—target spatial distance, radial velocity, azimuth angle, and radar cross section intensity—are extracted to construct a high-purity radar point cloud dataset. This simultaneous extraction of multi-dimensional scattering features enhances the feature representation capability of vehicle targets, providing high-precision radar feature data for subsequent multi-feature association matching and congested vehicle clustering fusion.

[0042] In one feasible embodiment, a compliant 93GHz roadside millimeter-wave radar is selected, with a transmission wavelength of... Equivalent radiated power With a range resolution of 0.2m, a velocity resolution of 0.12m / s, and an IP67 protection rating, it possesses excellent low-speed vehicle detection capabilities. Addressing the issues of high static clutter intensity and easy submersion of low-speed echoes in congested scenarios, a time-frequency joint analysis and multi-mechanism fusion clutter suppression scheme is employed to refine the radar's effective features. Specific implementation methods include the following: (1) Radar raw signal transformation and processing: The radar acquires intermediate frequency IQ complex sampling data in real time, and the number of sampling points in the range dimension. Number of sampling points in velocity dimension The time-frequency conversion is performed using a two-dimensional fast Fourier transform, and the calculation formula is as follows:

[0043] in: The radar time-domain IQ sampling signal is converted into a range-Doppler spectrum to finely characterize the micro-motion echo features of congested vehicles.

[0044] (2) Hybrid adaptive clutter suppression: Set the inter-frame differential time interval The formula for inter-frame difference calculation is:

[0045] in: This is the spectrum signal of the current frame. Historical frame spectral signals are used to remove static clutter from the road surface and guardrails; a zero-frequency notch filter is used to suppress ground DC interference. Traffic flow adaptive adjustment coefficient. The value range is [0.8, 1.4], and the adaptive threshold expression is:

[0046] in: The noise average of the clutter reference unit is used to increase the adjustment coefficient under congested conditions to prevent low-speed vehicle echoes from being filtered out by the threshold.

[0047] (3) Multidimensional scattering feature extraction: The radial velocity of the vehicle is solved based on the Doppler effect. The formula for calculating the radial velocity is as follows:

[0048] in: To determine the target's Doppler frequency shift, effective radar points with a signal-to-noise ratio greater than 10dB are selected. Simultaneously, four types of features—target range, azimuth, radial velocity, and RCS scattering intensity—are extracted to construct a high-purity radar point cloud dataset.

[0049] In terms of visual image processing, to address image degradation caused by lighting fluctuations, weather factors such as rain, fog, and dust, as well as noise and artifacts in real-world road conditions, multi-level joint preprocessing can be performed on the acquired images, including: Image dimensionality reduction is achieved by using weighted grayscale. Based on the photosensitive characteristics of the human eye, fixed weighting coefficients are set for the red, green, and blue channels to maximize the preservation of the image's light and dark texture features while reducing the computational load.

[0050] Two-dimensional Gaussian filtering is used to achieve smooth noise reduction, which removes salt-and-pepper noise and small specks while protecting the vehicle's edge contours from being blurred.

[0051] To address the issue of image blurring caused by atmospheric scattering under adverse weather conditions such as rain, fog, and dust, a dark channel calculation model is constructed based on the dark channel prior theory. By utilizing the inherent characteristics of low-brightness pixels in the local neighborhood of fog-free images, atmospheric transmittance is inferred in reverse, accurately eliminating interference noise caused by fog and dust, and restoring the true image quality.

[0052] Through the coordinated efforts of the aforementioned multi-level preprocessing algorithms, image degradation caused by uneven lighting, noise, and haze is corrected at multiple levels, maximizing the restoration of vehicle edge contours and body texture details.

[0053] In object detection, a lightweight SSD detection model with an embedded channel attention mechanism is constructed. Traditional SSD models lack feature extraction capabilities in scenarios with densely packed targets and occluded blind spots. This invention embeds a channel attention module into the SSD network, calculating the attention weights of feature channels through global average pooling and a multilayer perceptron, adaptively adjusting the feature weight distribution to enhance key vehicle texture features while suppressing interference from invalid background features. Simultaneously, the convolutional layers of the backbone network are simplified, reducing redundant convolutional units and effectively compressing the number of model parameters and computational overhead. This improved model integrates a lightweight structure and an attention weighting mechanism, optimizing feature extraction performance in complex traffic scenarios such as large vehicle occlusion and dense road congestion, significantly improving the detection accuracy of both clustered and occluded vehicles, while also meeting the real-time deployment requirements of edge-embedded devices. Based on the improved detection model after training, the image target is analyzed and reasoned. Multi-dimensional visual features such as pixel coordinates, body outline, vehicle type, and detection confidence of vehicle targets are stably extracted. A standardized visual target detection dataset is constructed. Compared with the traditional single visual sampling method, this invention integrates multiple features such as morphology, position, and probability to enrich the dimensions of visual representation. The feature data is standardized and structured and encapsulated to construct a standardized visual target detection dataset, which provides reliable visual data support for subsequent radar-visual heterogeneous feature association matching and congested vehicle clustering fusion.

[0054] The specific implementation method is as follows: (1) Multi-level joint image preprocessing: The camera acquires visible light color images of the road in real time, and adopts a multi-level preprocessing strategy of grayscale conversion, Gaussian noise reduction, and dark channel dehazing to optimize image quality and retain key vehicle texture information, including: Image grayscale conversion: This invention uses a weighted grayscale conversion method to complete image dimensionality reduction processing. Fixed weighting coefficients are set based on the light-sensing characteristics of the human eye to preserve the image's brightness and texture features to the maximum extent. The grayscale transformation calculation formula is as follows:

[0055] in: To output the pixel values ​​of the grayscale image; , , These represent the red, green, and blue pixel components of a color image; 0.299, 0.587, and 0.114 are standard optical empirical weights, corresponding to the human eye's visual perception sensitivity to the red, green, and blue bands, respectively, avoiding the loss of texture details caused by simple mean grayscale.

[0056] Gaussian noise reduction: The expression for the Gaussian filter kernel function is as follows:

[0057] in: This is the Gaussian filter kernel function; These are local relative coordinates of the pixels; This is the Gaussian smoothing coefficient, used to adjust the degree of noise smoothing.

[0058] Dark channel dehazing optimization: The dark channel expression is as follows:

[0059] in: These are the pixel values ​​of the dark channel of the image. These are the single-channel pixel components corresponding to the original color image; In pixels A local pixel neighborhood window centered on the pixel.

[0060] By employing a three-level preprocessing algorithm that combines grayscale conversion, Gaussian noise reduction, and dark channel dehazing, the image degradation caused by uneven lighting, noise, and haze is corrected in multiple ways, maximizing the restoration of vehicle edge contours and body texture details, and improving image quality under complex traffic conditions.

[0061] (2) An improved single-shot multi-frame detector (SSD) lightweight detection model with embedded attention mechanism: The SSD model relies on the anchor frame mechanism to locate and classify vehicle targets. It corrects the detection box position by predicting the anchor frame offset. Its bounding box regression calculation formula is as follows:

[0062] in: To predict the coordinates and width / height parameters of the bounding box; These are the original prior anchor frame parameters; , , This is the offset correction amount obtained through network autonomous learning.

[0063] This invention embeds a channel attention module to adaptively adjust the feature weight distribution. The attention weight calculation formula is as follows:

[0064] in: These are the feature attention weight coefficients; Input feature maps into the network; This is a global average pooling operation; It is a multilayer perceptron; This is the Sigmoid activation function, used to perform weight normalization.

[0065] Meanwhile, this model simplifies the convolutional layers of the backbone network, reduces redundant convolutional units, effectively compresses the number of model parameters and computational overhead, and meets the real-time deployment requirements of roadside embedded devices.

[0066] (3) Visual texture feature aggregation and standardized dataset construction: Based on the improved detection model after training, the image target is reasoned and analyzed to stably extract multi-dimensional visual features of the vehicle, including image pixel coordinates, body outline, vehicle type, and detection confidence.

[0067] After extracting radar scattering features and visual texture features separately, multi-feature collaborative fusion clustering processing is performed, including: (1) Based on the external parameters obtained from the previous calibration, coordinate transformation is completed, and the three-dimensional radar point cloud is projected and mapped to the two-dimensional image pixel coordinate system to achieve the unification of heterogeneous sensors in the spatial dimension. On this basis, the association and binding relationship between radar point cloud and visual detection target is established based on feature matching criteria, and the radar motion features and visual texture features of vehicles of the same origin are accurately matched to construct a one-to-one fusion feature dataset, thereby eliminating the problems of spatial misalignment and matching disorder of heterogeneous features.

[0068] (2) To achieve intelligent adaptive adjustment of clustering parameters, the density of traffic flow is characterized by the number of effective radar point clouds within a unit area, and two types of traffic conditions are quantitatively divided: smooth traffic flow and congested traffic flow. The traffic flow density determination formula is as follows:

[0069] In the formula: Point cloud density per unit area; This represents the number of valid radar points within the area. To define the area for testing.

[0070] (3) This method abandons the drawback of setting fixed hyperparameters in the traditional DBSCAN clustering algorithm. It dynamically adjusts the core parameters of clustering based on the traffic density discrimination results. Under congested conditions, vehicles are densely distributed and point clouds are heavily stacked. In this case, the clustering neighborhood radius is reduced and the core point threshold is increased to avoid the problem of adjacent vehicles mixing and sticking together. Under smooth conditions, vehicles are discretely distributed and point clouds are sparse. The clustering neighborhood radius is appropriately expanded and the core point requirement is reduced to ensure that discrete vehicles with the same origin are clustered normally. Through this intelligent parameter tuning method that adapts to different traffic conditions, the generalization ability of the algorithm under different traffic scenarios is significantly improved.

[0071] The adaptive decision-making logic formula is as follows:

[0072] in: The radius of the cluster neighborhood; The minimum number of core points in the neighborhood; The threshold for determining traffic flow density.

[0073] (4) To solve the technical problems of noise interference and multiple vehicle clustering in dense traffic flow, a dual constraint discrimination condition is constructed from two dimensions: motion characteristics and visual morphology. The speed deviation constraint ensures the consistency of target motion within the cluster, that is, the deviation between the motion speed of each point in the cluster and the average speed of the cluster does not exceed the preset speed threshold; combined with the visual detection box intersection-union constraint, the homology of vehicle contours is verified, that is, the visual detection box intersection-union ratio of each target in the same cluster is not lower than the preset contour overlap threshold. Through the dual constraint condition, non-homologous interference points with speed anomalies and spatial displacements are accurately eliminated, the feature purification of the cluster is completed, and the problems of dense vehicle adhesion and misclassification are effectively improved.

[0074] The algorithm constraint determination formula is as follows:

[0075] in: This is the speed deviation value; The velocity is the velocity of a single point. The average speed of the cluster; The speed constraint threshold; The visual contour intersection-union ratio function; The threshold for contour overlap constraint.

[0076] After high-precision clustering and purification, a weighted fusion algorithm is used to couple radar and visual dual-source features. Weights are allocated differently based on the sensor detection confidence levels to compensate for the shortcomings of single-sensor detection and to fit and optimize the multi-dimensional state parameters of the vehicle. Finally, the fitting solution obtains high-precision perception parameters such as the vehicle's precise location, speed, vehicle type, vehicle size, and lane, enabling refined intelligent perception of vehicle targets in complex urban road congestion scenarios. The formula for the weighted fusion algorithm is as follows:

[0077] in: For vehicle fusion feature parameters; This represents the single-point motion characteristics of radar. Visual texture contour features; , These are the feature weighting coefficients for the two types of sensors, respectively.

[0078] (5) Finally, the precise location, speed, vehicle type, body size, lane and other high-precision perception parameters of the vehicle are obtained by fitting and solving, so as to realize the refined and robust intelligent perception of vehicle targets in complex urban road congestion scenarios.

[0079] After completing the clustering and fusion of vehicle targets, multi-feature joint multi-target tracking is performed, including: Vehicle motion state modeling: A constant turning rate and velocity model is used to characterize the vehicle's driving state. A vehicle state vector is constructed, containing planar coordinates, linear velocity, heading angle, and yaw rate. The state vector encompasses multi-dimensional information such as vehicle position, speed, and yaw attitude, and can adapt to complex nonlinear motion conditions such as vehicle start-stop, low-speed slow-movement, and lane changing. It is represented as:

[0080] By combining discrete-time sampling sequences, a discrete motion equation for the vehicle is established. Based on this dynamic model, a Kalman filter iterative framework is constructed, comprising two stages: state prediction and observation update. The Kalman filter iterative update formula is as follows:

[0081] Setting process noise covariance Observation noise covariance It conforms to the low-noise detection characteristics of the roadside.

[0082] Multi-feature cost association: Based on the vehicle motion state of the filtered output, combined with visual contour and texture information, the position deviation cost, velocity deviation cost, and visual appearance similarity cost of the target between frames are comprehensively calculated. A multi-feature joint cost function is constructed to quantify the matching difficulty between different targets. The multi-dimensional comprehensive cost function is expressed as follows:

[0083] Weighting coefficient , , A two-dimensional cost matrix is ​​constructed based on the pairwise target matching cost. The Hungarian algorithm is used to solve for the optimal matching relationship with the global minimum matching cost as the optimization objective. This ensures a unique match between the trajectory and the detected target, avoids the cross-matching mismatch problem caused by densely connected vehicles, and improves the stability of inter-frame data association in complex traffic environments. The global matching constraint is as follows:

[0084] Sensor-coordinated occlusion compensation: To address the issues of single-sensor target misses and feature loss caused by congestion and occlusion, an occlusion tolerance threshold is set to construct a trajectory compensation mechanism. The maximum occlusion tolerance frame count can be set to 4 frames. During the occlusion period, the target position is continuously deduced based on the filtering equation to avoid trajectory interruption.

[0085] Trajectory lifecycle management: Set the number of consecutive valid feature accumulation frames as the new trajectory determination condition. Two consecutive valid detection frames determine the new trajectory and assign a unique ID. Set the number of consecutive disappearance frames as the disappearance determination condition. If there is no valid feature response for a long time, the target is determined to have left the field of view and trajectory updates are terminated. For example, if there is no response for 6 consecutive frames, the target is determined to have left the field of view, trajectory updates are terminated, and false noise trajectories are removed.

[0086] Through the hierarchical tracking logic of dynamic modeling, multi-feature association, occlusion compensation, and lifecycle management, stable and continuous tracking of multiple vehicles in complex congested and occluded scenarios is achieved. High-quality vehicle trajectories with unique identifiers and no frequent jumps are output, providing reliable time-series data support for subsequent traffic flow analysis and driving behavior assessment. After completing multi-target tracking, a congestion-adaptive dual-mode traffic flow counting is implemented. A unique identifier constraint is established based on the unique tracking identifier output from multi-target tracking. When a vehicle's trajectory crosses the boundary of the statistical zone and the tracking identifier is not repeatedly recorded, it is considered a single valid passage event. Valid traffic flow is filtered based on the spatiotemporal variation characteristics of the trajectory, and duplicate sampling data of vehicles traveling in the same direction is eliminated using the unique identifier constraint, ensuring the rigor of traffic flow statistics under smooth road conditions. To address the challenges of vehicle stacking and adhesion, and the difficulty of cross-line statistics due to stationary trajectories under congested conditions, a congestion counting model based on time-dwelling characteristics is established. A vehicle dwell time threshold is set; when a vehicle's continuous dwell time in the intermediate statistical zone reaches this threshold, it is considered a queued vehicle and counted. Simultaneously, a periodic deduplication constraint is added, limiting the same tracking identifier to only one count within a single congestion cycle, avoiding the problem of high-frequency duplicate counting in low-speed congestion scenarios from an algorithmic perspective. To address the issues of limited detection field of view and missed detection of distant vehicles due to queue overflow in extreme congestion, a dynamic fitting and correction mechanism for traffic flow is constructed. A regional saturation capacity threshold is set. When the number of vehicles in the statistical area in real time exceeds this threshold, the traffic flow is considered to have reached queue overflow saturation. A linear fitting function is constructed based on the upstream traffic flow entry rate, and the scale of queued vehicles outside the field of view is extrapolated by combining the time accumulation, thus correcting the statistical bias under extreme congestion conditions. Finally, the statistical results of multiple modes are integrated, and multi-dimensional traffic parameters such as lane-level traffic flow, road segment comprehensive flow, average vehicle speed in the section, road traffic congestion density, and vehicle type distribution ratio are extracted. The original statistical data is then subjected to noise reduction, fitting, and normalization processing to generate standardized structured traffic monitoring reports.

[0087] In one feasible embodiment, the sensing area is divided into an upstream input area, an intermediate statistics area, and a downstream output area, with each segment having a longitudinal length of 50m. The counting logic is adaptively switched based on vehicle speed, using the real-time vehicle speed obtained from fusion detection as the classification criterion. A traffic flow classification speed threshold is set, and the traffic flow condition is classified by comparing the vehicle fusion speed with the classification threshold. When the vehicle fusion speed is greater than the threshold, it is determined to be a smooth, free-flowing traffic flow; when the vehicle fusion speed is greater than or equal to zero but not greater than the threshold, it is determined to be a low-speed, congested traffic flow. This discrimination method accurately distinguishes between free-flowing and low-speed queuing traffic flows based on vehicle motion characteristics, providing a basis for the adaptive switching of subsequent counting modes.

[0088] Set traffic flow classification speed thresholds The determination formula is as follows:

[0089] Smooth crossing count: Based on the unique tracking ID, if a vehicle's trajectory crosses the statistical boundary and there are no duplicate records, it is determined to be a valid passage.

[0090] Congestion Dwell Count: Dwell Time Threshold , Judgment condition:

[0091] The same ID is counted only once within a single congestion cycle to avoid duplicate counting.

[0092] Queue overflow compensation: Regional saturation capacity threshold Traffic density correction factor The supplementary calculation formula is as follows:

[0093] The number of vehicles queuing in remote blind spots is extrapolated using integral calculations to correct statistical biases in extreme congestion.

[0094] Data output: Integrates lane flow, average vehicle speed, congestion density, and vehicle type ratio to generate standardized traffic monitoring reports.

[0095] After completing the traffic flow counting, multi-feature joint anomaly verification is performed. A dual-source confidence fusion judgment model is constructed, combining radar detection confidence and visual detection confidence. The two are weighted and summed to obtain the target's comprehensive confidence. The weighting coefficients are adaptively fine-tuned according to the traffic congestion status. Under normal traffic conditions, the weights of the two are equal to maintain balance. Under congested conditions, the weight coefficient of radar detection confidence is appropriately increased to resist visual occlusion interference. Under smooth traffic conditions, the visual weight is appropriately increased to enhance feature recognition accuracy, achieving complementary advantages of the two sensors. A target retention confidence threshold is set, and the comprehensive confidence is used to distinguish between valid targets and false interference targets. The algorithm adopts a relaxed retention logic, which determines a target as valid as long as it meets the high confidence detection condition of a single sensor, minimizing the omission of real targets in congested and occluded scenarios. Noisy false targets with low comprehensive confidence and incomplete features are filtered out to reduce detection disturbances caused by environmental stray interference. Redundant data cleaning is performed on the selected and retained valid targets to remove invalid data with duplicate sampling, feature distortion, and abnormal motion. By combining multi-target tracking trajectory constraints to verify the rationality of target motion, abnormal discrete data with sudden velocity changes and position jumps are eliminated. Through dual purification at both the target and data levels, the false detection and misdetection rates are effectively reduced, improving the data quality of the entire detection and tracking process and providing high-purity, reliable data for vehicle tracking and traffic counting.

[0096] In one feasible embodiment, the dual-source confidence fusion determination is performed using the following formula:

[0097] Normal operating conditions The congestion condition has been adjusted to , This enhances the radar's ability to resist obstruction.

[0098] Abnormal target filtering and removal: Set retention confidence threshold Filtering and judgment rules:

[0099] Data optimization and correction: Abnormal discrete data with velocity jumps greater than 2.0 m / s and position jumps greater than 1.0 m are removed to ensure data purity.

[0100] Meanwhile, this invention also discloses a multi-feature collaborative enhancement congestion traffic flow radar-visual fusion detection system for implementing the above-mentioned method. It includes a coaxially mounted millimeter-wave radar and camera as an acquisition module, an embedded edge computing terminal, and a data storage and transmission module. The acquisition module simultaneously acquires road radar point cloud and visual image data. The edge computing terminal is equipped with a lightweight fusion detection algorithm to complete multi-source data processing, vehicle perception, and traffic flow statistics, suitable for traffic monitoring of normally congested urban road sections. The embedded edge computing terminal is equipped with the aforementioned multi-feature collaborative enhancement fusion detection method, internally integrating a data acquisition and hierarchical spatiotemporal synchronous registration module, a radar scattering feature extraction module, a visual texture feature extraction module, a multi-feature collaborative fusion clustering module, a multi-feature joint multi-target tracking module, a congestion adaptive dual-mode traffic flow counting module, and a multi-feature joint anomaly verification module. The modules interact with each other through standardized data interfaces, forming a complete processing link from front-end data acquisition to back-end detection result output. The edge computing terminal adopts a multi-core processor architecture, possessing strong parallel computing capabilities while maintaining low power consumption, adapting to the heat dissipation conditions and power supply constraints of roadside enclosures. The terminal integrates a hardware level-triggered synchronization interface. The radar's sampling clock serves as a global time reference. The camera's acquisition time is synchronized through a hardware trigger signal, keeping the time synchronization error of the two sensors within a preset allowable range and eliminating time misalignment caused by asynchronous acquisition.

[0101] At the data stream level, the raw intermediate frequency (IF) IQ signal from the radar is transmitted to the radar scattering feature extraction module of the edge computing terminal via a high-speed data interface. This module performs time-frequency transformation and hybrid clutter suppression on the radar signal, extracting multi-dimensional scattering features of the vehicle target, such as distance, radial velocity, azimuth angle, and radar cross-section intensity, to generate a structured radar point cloud dataset. Simultaneously, image data acquired by the camera is transmitted to the visual texture feature extraction module via an image transmission interface. This module performs multi-level preprocessing on the image and uses a lightweight target detection model with an embedded attention mechanism for inference and parsing, extracting visual texture features such as pixel coordinates, vehicle outline, vehicle type, and detection confidence, to generate a standardized visual target detection dataset. The two data streams are processed in parallel within the edge computing terminal, and then integrated by a subsequent fusion module.

[0102] The data storage and transmission module is responsible for storing intermediate data and transmitting final detection results during processing. This module includes a local solid-state storage unit and a remote communication unit. The local storage unit archives the detection results in a structured manner according to timestamps and lane numbers, forming traceable historical detection records. The remote communication unit uses a common network communication protocol to upload multi-dimensional traffic parameters such as lane-level traffic flow, road segment comprehensive traffic flow, average vehicle speed in the section, road traffic congestion density, and vehicle type distribution ratio to the superior traffic management platform or cloud data center in real time. The communication unit has breakpoint resume and data caching functions. When the network is unstable, it automatically caches the detection data and completes the retransmission after the network is restored, ensuring the integrity and reliability of the data.

[0103] Regarding the inter-module collaboration, the synchronous data output from the data acquisition and hierarchical spatiotemporal synchronization registration module is fed into the radar scattering feature extraction module and the visual texture feature extraction module, respectively. The feature data output from both modules are then fed into the multi-feature collaborative fusion clustering module. This module performs correlation mapping of heterogeneous features, adaptive clustering segmentation, and weighted fusion fitting, outputting vehicle state parameters. The output from the multi-feature collaborative fusion clustering module is fed into the multi-feature joint multi-target tracking module. This module performs motion state estimation, inter-frame correlation matching, and occlusion compensation, outputting continuous and stable vehicle trajectories. The output from the multi-target tracking module is fed into the congestion adaptive dual-mode traffic flow counting module. This module performs traffic flow state discrimination, counting mode switching, and queue overflow compensation, outputting traffic statistics parameters. The multi-feature joint anomaly verification module performs joint confidence judgment and data purification on the output data from each of the above stages, ensuring the purity and reliability of the detection results from an overall perspective.

[0104] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative and not exhaustive. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A multi-feature collaborative enhancement method for congested traffic flow radar-visual fusion detection, characterized in that, Includes the following steps: Acquire raw echo signals from millimeter-wave radar and image data captured by a camera; A hierarchical spatiotemporal synchronization architecture, which includes hardware time hard synchronization and dynamic spatial deviation correction, is adopted to synchronize and align the radar signal and the image in both time and spatial dimensions. The synchronized radar signal is subjected to time-frequency transformation, and a hybrid clutter suppression algorithm that includes dynamic adjustment of the suppression threshold based on traffic density is used to remove static environmental clutter and extract the radar scattering characteristics of vehicle targets. The synchronized image is processed using a lightweight target detection model with an embedded attention mechanism to extract the visual texture features of the vehicle target. The radar point cloud is projected onto the image coordinate system, and a clustering algorithm with adaptive adjustment of clustering neighborhood parameters according to traffic flow density is used to segment the fused features. The segmentation results are purified by speed consistency constraints and contour overlap constraints, and vehicle state parameters are fitted by weighted fusion. Motion state estimation is performed based on the vehicle kinematics model. Inter-frame target association matching is performed using a cost function that integrates position features, velocity features, and appearance features. Trajectory prediction compensation is set under occlusion conditions to achieve continuous tracking of multiple targets. The system determines the traffic congestion status based on the real-time vehicle speed, adaptively switches between the smooth cross-line counting mode and the congestion dwell counting mode, and performs supplementary vehicle counting for queue overflow scenarios. The confidence scores of radar detection and visual detection are adaptively weighted and fused to obtain a comprehensive confidence score. Valid targets are selected and false targets are eliminated based on the comprehensive confidence score threshold.

2. The method according to claim 1, characterized in that, The hardware time synchronization is as follows: using the radar sampling clock as the global time reference, a hardware level-triggered synchronization mechanism is adopted to constrain the time error between the radar sampling time and the camera acquisition time within a preset allowable range.

3. The method according to claim 1, characterized in that, The dynamic spatial deviation correction is as follows: continuously extract stable feature points of the road at preset time intervals, and use the least squares method to iteratively optimize and update the spatial transformation matrix to compensate for the cumulative registration error caused by environmental disturbances.

4. The method according to claim 1, characterized in that, The hybrid clutter suppression algorithm includes: using inter-frame differential operation to remove static background clutter, using zero-frequency notch filtering to suppress ground DC low-frequency interference, and dynamically adjusting the constant false alarm threshold based on real-time traffic density.

5. The method according to claim 1, characterized in that, The lightweight object detection model with embedded attention mechanism is as follows: a channel attention module is embedded in the SSD object detection network, and the channel attention module calculates the attention weights of the feature channels through global average pooling and multilayer perceptron; the backbone network of the object detection model is simplified and compressed through convolutional layers.

6. The method according to claim 1, characterized in that, The adaptive adjustment of the clustering neighborhood parameters with traffic flow density is specifically as follows: traffic flow density is characterized by the number of effective radar point clouds within a unit area; under congested conditions, the neighborhood radius of the DBSCAN clustering algorithm is reduced and the minimum core point threshold is increased; under smooth conditions, the neighborhood radius is expanded and the minimum core point threshold is decreased; the speed consistency constraint is that the deviation between the movement speed of each point in the cluster and the average speed of the cluster does not exceed a preset threshold; the contour overlap constraint is that the intersection-union ratio of the visual detection boxes of each target in the same cluster is not lower than a preset threshold.

7. The method according to claim 1, characterized in that, The vehicle kinematics model is a constant turning rate and speed model; the cost function is composed of a weighted sum of position deviation cost, speed deviation cost and visual appearance similarity cost; the inter-frame target association matching uses the Hungarian algorithm to solve for the optimal matching relationship with the global minimum matching cost as the optimization objective.

8. The method according to claim 1, characterized in that, The trajectory prediction compensation under the occlusion state is as follows: when the detection signal of a single sensor is missing, the vehicle position is continuously predicted iteratively by relying on the redundant feature information of heterogeneous sensors and Kalman filtering; and, the number of consecutive effective feature accumulation frames is set as the condition for the new trajectory, and the number of consecutive disappearing frames is set as the condition for the trajectory to disappear, so as to realize trajectory life cycle management.

9. The method according to claim 1, characterized in that, The smooth crossing counting mode is as follows: traffic flow statistics are completed by using trajectory crossing discrimination. When a vehicle's trajectory crosses the boundary of the statistical area and the tracking identifier is not recorded repeatedly, it is determined as a single valid passage event. The congestion dwelling counting mode is as follows: queued vehicles are determined by the duration of continuous dwelling of vehicles in the statistical area, and the same tracking identifier is only counted once in a single congestion cycle. The queue overflow supplementary counting is as follows: a regional saturation capacity threshold is set. When the number of vehicles in the statistical area in real time is greater than the threshold, the scale of queued vehicles outside the field of view is extra-counted based on the upstream traffic flow entry rate by constructing a fitting function.

10. A multi-feature collaborative enhancement system for traffic congestion detection using radar-visual fusion, characterized in that, The detection method according to any one of claims 1 to 9 includes a coaxially mounted millimeter-wave radar and camera acquisition module, an embedded edge computing terminal, and a data storage and transmission module; The embedded edge computing terminal includes: The data acquisition and hierarchical spatiotemporal synchronization registration module is used to acquire radar signals and image data and perform synchronization and alignment using a hierarchical spatiotemporal synchronization architecture. The radar scattering feature extraction module is used to perform time-frequency transformation on radar signals and extract radar scattering features using a hybrid clutter suppression algorithm; The visual texture feature extraction module is used to extract visual texture features using a lightweight object detection model with an embedded attention mechanism. A multi-feature collaborative fusion clustering module is used for projection mapping, adaptive clustering, dual-constraint purification, and weighted fusion fitting of vehicle state parameters; A multi-feature joint multi-target tracking module is used for motion state estimation, multi-feature cost association matching, and occlusion compensation tracking; The congestion adaptive dual-mode traffic counting module is used for traffic flow status identification, adaptive switching of counting modes, and queue overflow supplementary counting. A multi-feature joint anomaly verification module is used for adaptive fusion of dual-source confidence and removal of false targets.

Images