A big data-based vehicle driving data analysis system and method

Abstract

The present application relates to the field of vehicle monitoring, in particular to a vehicle driving data analysis system and method based on big data, comprising: a radar acquisition module, a signal identification module, a vehicle positioning module, a motion speed measurement module and a trajectory snapshot module, the radar acquisition module is used for building a radar acquisition platform and storing double-channel spectrum, the signal identification module is used for determining a vehicle target area, the vehicle positioning module is used for constructing a vehicle contour model, the motion speed measurement module is used for measuring the speed of a target vehicle, and the trajectory snapshot module is used for outputting a safe speed limit and snapshotting an overspeed target, the present application can improve the accuracy of vehicle detection and snapshotting, improve the accuracy and reliability of target identification, filter out radar noise, improve the signal-to-noise ratio, realize long-distance vehicle sensing, road trajectory modeling and road detection, avoid vehicle overspeed, assist in optimizing the driving path of the vehicle and improve the road traffic management capability.

Description

Technical Field

[0001] This invention relates to the field of vehicle monitoring, specifically to a vehicle driving data analysis system and method based on big data. Background Technology

[0002] Speed ​​radar is a device that uses the Doppler effect to measure vehicle speed. It transmits radio waves and receives the reflected echoes, calculating the frequency shift between the transmitted and reflected waves to obtain the vehicle's instantaneous speed. Speed ​​radar is less affected by environmental factors such as lighting and weather, and can detect vehicle speed over long distances, achieving the goals of road traffic awareness and refined management.

[0003] Because road radar monitoring points are sparsely and unevenly distributed, speed radar has a large measurement error on roads and it is difficult to cover a wide area of ​​the road surface. Commonly used conical radar and flat panel radar have complex detection processes, low false negative rates, slow data processing speeds, and are easily affected by other signals, resulting in noise. The threshold setting is also difficult to control, making it difficult to guarantee the accuracy of target detection.

[0004] Furthermore, due to the different speed limits for different roads and vehicle types, existing radar capture devices cannot distinguish between large and small vehicles on the highway, cannot modify the speed limit standards of road sections according to the vehicles and road environment, and cannot achieve full-area vehicle location perception, thus failing to continuously track vehicles across road sections. Summary of the Invention

[0005] The purpose of this invention is to provide a vehicle driving data analysis system and method based on big data to solve the problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a vehicle driving data analysis system based on big data, comprising: a radar acquisition module, a signal recognition module, a vehicle positioning module, a motion speed measurement module, and a trajectory capture module; The radar acquisition module is used to build a radar acquisition platform in the road, modulate radio frequency signals, change the elevation angle of the radar surface to adjust adjacent detection domains, receive the feedback signals, process them and store the echoes as I-channel and Q-channel spectra, and record the signal frequency and phase information respectively. The signal recognition module is used to extract intermediate frequency signal features in the time domain, construct a feature space, amplify the processed signal at the intermediate frequency, perform AD sampling after amplification, perform FFT fast Fourier transform on the I and Q data, convert the discrete time signals of the I and Q channels into frequency domain signals, set signal strength thresholds for large and small vehicles, count the number of frames greater than the threshold, determine the vehicle length based on the frame count results, predict the physical location of the scattering center, and determine the vehicle target area. The vehicle positioning module is used to acquire the two-dimensional Doppler spectrum of the radar on the vehicle target area, and to track and extract the point cloud of the vehicle in the target area through CFAR peak detection and DOA direction detection to obtain the vehicle point cloud. Based on the vehicle length and point cloud distribution characteristics, the vehicle type is classified and a vehicle outline model is constructed. The motion speed measurement module is used to illuminate the vehicle outline model, measure the vehicle spectrum, interpolate the Doppler frequency and compensate for acceleration, eliminate the difference frequency Doppler spectrum of the signal through phase cancellation, accumulate the difference frequency Doppler spectrum, and obtain the target vehicle speed information. The trajectory capture module is used to collect the driving trajectories of each vehicle on the road, construct a vehicle trajectory dataset, build a BP neural network, output real-time safe speed limits, and use a camera to capture the target when the vehicle speed exceeds the speed limit corresponding to the vehicle type.

[0007] Furthermore, the radar acquisition module includes: a signal modulation unit and a dual-channel spectrum unit; The signal modulation unit is used to output an intermediate frequency signal through a piezoelectric oscillator using a 24GHz or 77GHz frequency-modulated continuous wave radar. After receiving the echo, it constructs a semantic segmentation network, removes non-road semantic interference, suppresses clutter in the target echo, extracts feature parameters from the Doppler spectral domain, divides the road area, and filters out roadside interference. The dual-channel spectral unit is used to calculate the cross-correlation coefficient of the dominant components of the signal, identify the radar signal, reconstruct and denoise the signal, and obtain two baseband signals, I and Q, through orthogonal demodulation, which represent the amplitude and phase information of the signal, respectively. The I / Q signals are stored in complex form to form a dual-channel output.

[0008] Furthermore, the signal recognition module includes: a noise suppression unit, a frequency domain analysis unit, and a target localization unit; The noise suppression unit is used to determine the background subspace of each channel through principal component analysis, construct a projection matrix, project the signal onto the orthogonal background subspace, generate a masking matrix to weight the mixed signal components in the high-dimensional space, multiply each component in the feature space by the component weight, and suppress the background signal. The frequency domain analysis unit is used to delay and add frequency domain signals to improve the sliding window energy ratio and the peak height of the detected signal, analyze the frequency domain signal, reconstruct the loss function based on the echo intensity weighting, establish the relationship between frequency domain characteristics and vehicle scattering structure, input the scattering predictor, and measure the position of the scattering center. The target localization unit is used to scan the prior target area of ​​the previous frame, guide the radar to select target boxes with waveform numbers during the scanning process, monitor the vehicle position and adjust the position deviation, encode the target boxes into a high-dimensional space, use the feature distance as the temporal association cost, and the triplet loss encoder learns the geometric similarity of the matching boxes to achieve three-dimensional vehicle target tracking.

[0009] Furthermore, the vehicle positioning module includes: a target scanning unit and a contour modeling unit; The target scanning unit is used to extract the signal features of the transmitted signal as tags to optimize the dual-channel radio frequency, constrain the input of the received signal through the interference suppression autoencoder, and calculate the azimuth and elevation angles of the target based on the two-dimensional Doppler spectrum to obtain the three-dimensional spatial coordinates of each detection point, forming a vehicle point cloud. The contour modeling unit is used to aggregate the point clouds of the same vehicle onto the same time axis, cluster and segment the vehicles, calculate the two-dimensional and three-dimensional clustering errors based on the prior vehicle length during clustering, adjust the clustering radius and minimum number of points, classify the vehicles according to the distribution range of the point cloud, and construct the vehicle contour model based on the classification results.

[0010] Furthermore, the motion speed measurement module includes: a data compensation unit and a speed measurement unit; The data compensation unit is used to perform FFT transformation on the frequency domain signal at the same distance, refine the spectrum by interpolation, suppress the difference frequency component by double pulse cancellation, improve the signal-to-noise ratio by coherently accumulating pulse trains, and obtain the difference frequency Doppler spectrum. The velocity measurement unit is used to calculate the target radial velocity according to the Doppler frequency corresponding to the peak value of the difference frequency Doppler spectrum, and to determine the real-time velocity of the target vehicle by converting the frequency and velocity.

[0011] Furthermore, the trajectory capture module includes: a trajectory modeling unit, a driving monitoring unit, and a dynamic speed limiting unit; The trajectory modeling unit is used to transform the coordinates of adjacent radar data, determine the vehicle trajectory, calculate the trajectory similarity through motion sequences, filter out scattering and split trajectories, identify the trajectory of the same target, supplement missing trajectory points, perform feature analysis on vehicle trajectory data, extract continuous trajectories of different targets according to time feature patterns, and formulate status labels. The driving monitoring unit is used to construct a trajectory network. The nodes are composed of radar collection points, and the directed edges are composed of clustered GPS trajectories. It includes all driving paths of the vehicle in the city. When the vehicle passes through an adjacent node, the probability of each path is predicted, and the path with the highest probability is selected as the vehicle's spatial trajectory. The dynamic speed limit unit is used to create a training set based on historical traffic flow data, vehicle operation records, accident records, and road markings. The mean squared error is used as the loss function to train a BP neural network. The unit outputs a safe speed limit based on vehicle trajectory, number of vehicles, vehicle type, and vehicle speed as input. When the vehicle speed exceeds the safe speed limit, the unit triggers the camera to capture the image.

[0012] A vehicle driving data analysis method based on big data includes the following steps: Step S1. Build a radar acquisition platform, modulate the radio frequency signal for transmission, receive the echo signal, process it, and store the echo as I-channel and Q-channel spectra, recording the signal frequency and phase information respectively; Step S2. Extract the intermediate frequency signal features, construct the feature space, amplify the processed signal and perform AD sampling, perform FFT transformation on the two data streams to convert the discrete-time signal to the frequency domain signal, set the signal strength threshold, count the number of frames greater than the threshold, determine the vehicle length based on the frame count results, predict the physical location of the scattering center, and determine the target area. Step S3. Obtain the two-dimensional Doppler spectrum on the target area, and extract the point cloud of the vehicle in the target area by CFAR peak detection and DOA direction detection to obtain the vehicle point cloud. Classify the vehicle type according to the vehicle length and point cloud distribution characteristics, and construct the vehicle outline model. Step S4. Illuminate the vehicle contour model with radar, measure the vehicle spectrum, interpolate the Doppler frequency and compensate for acceleration, eliminate the difference frequency Doppler spectrum of the signal through phase cancellation, accumulate the difference frequency Doppler spectrum, and obtain the target vehicle speed information. Step S5. Collect the driving trajectories of each vehicle on the road, construct a vehicle trajectory dataset, train a neural network, output real-time safe speed limits, and use a camera to capture images of the target when the vehicle speed exceeds the speed limit corresponding to the vehicle type.

[0013] Furthermore, step S1 includes: Step S11. Using a 24GHz or 77GHz frequency-modulated continuous wave radar, the intermediate frequency signal is output through a piezoelectric oscillator. After receiving the echo, a semantic segmentation network is constructed to remove non-road semantic interference, suppress clutter in the target echo, extract feature parameters from the Doppler spectral domain, divide the road area, and filter out roadside interference. Step S12 calculates the cross-correlation coefficient of the dominant components of the signal, identifies the radar signal, reconstructs and denoises the signal, and obtains two baseband signals, I and Q, through orthogonal demodulation, which represent the amplitude and phase information of the signal, respectively. The I / Q signals are stored in complex form to form a dual-channel output.

[0014] Furthermore, step S2 includes: Step S21. Determine the background subspace of each channel through principal component analysis, construct a projection matrix, project the signal onto the orthogonal background subspace, generate a masking matrix to weight the mixed signal components in the high-dimensional space, multiply each component in the feature space by the component weight, and suppress the background signal. Step S22. Delay and sum the frequency domain signals to increase the sliding window energy ratio and the peak height of the detection signal. Analyze the frequency domain signal, reconstruct the loss function based on the echo intensity weight, establish the relationship between the frequency domain characteristics and the vehicle scattering structure, input the data into the scattering predictor, and measure the position of the scattering center. Step S23. Scan the prior target region of the previous frame, guide the radar to select target boxes with waveform numbers during the scanning process, monitor the vehicle position and adjust the position deviation, encode the target boxes into a high-dimensional space, use the feature distance as the temporal association cost, and use the triplet loss encoder to learn the geometric similarity of the matching boxes to achieve three-dimensional vehicle target tracking.

[0015] Furthermore, step S3 includes: Step S31. Extract the signal features of the transmitted signal as tags to optimize the dual-channel radio frequency. Constrain the input of the received signal by the interference suppression autoencoder. Calculate the azimuth and elevation angles of the target based on the two-dimensional Doppler spectrum to obtain the three-dimensional spatial coordinates of each detection point and form a vehicle point cloud. Step S32. Aggregate the point clouds of the same vehicle to the same time axis, cluster and segment the vehicles. During clustering, calculate the two-dimensional and three-dimensional clustering errors based on the prior vehicle length, adjust the clustering radius and minimum number of points, classify the vehicles according to the distribution range of the point cloud, and construct the vehicle outline model based on the classification results.

[0016] Furthermore, step S4 includes: Step S41. Perform FFT transformation on the frequency domain signal at the same distance, refine the spectrum by interpolation, suppress the difference frequency component by double pulse cancellation, improve the signal-to-noise ratio by coherently accumulating pulse trains, and obtain the difference frequency Doppler spectrum; Step S42. Based on the Doppler frequency corresponding to the peak value of the difference frequency Doppler spectrum, calculate the radial velocity of the target vehicle according to the Doppler principle, convert the frequency and velocity, and determine the real-time velocity of the target vehicle.

[0017] Furthermore, step S5 includes: Step S51. Transform the coordinates of adjacent radar data to determine the vehicle trajectory, calculate the trajectory similarity through motion sequence, filter out scattering split trajectories, identify the trajectory of the same target, supplement the missing trajectory points, perform feature analysis on the vehicle trajectory data, extract the continuous trajectories of different targets according to the time feature pattern, and formulate status labels. Step S52. Construct a trajectory network. The nodes are composed of radar collection points, and the directed edges are composed of clustered GPS trajectories. It includes all driving paths of the vehicle in the city. When the vehicle passes through an adjacent node, predict the probability of each path and select the path with the highest probability as the vehicle's spatial trajectory. Step S53. Create a training set based on historical traffic flow data, vehicle operation records, accident records, and road markings. Train a BP neural network using mean squared error as the loss function. Analyze vehicle trajectory, number of vehicles, vehicle type, and vehicle speed as inputs to output a safe speed limit. When a vehicle's speed exceeds the safe speed limit, trigger the camera to capture the image.

[0018] Compared with the prior art, the beneficial effects achieved by the present invention are: 1. This invention records signal frequency and phase information by setting up a radar acquisition platform on the road, analyzes the frequency domain signals of the I and Q channels, and determines the vehicle target area. This can improve signal diversity, remove non-road interference, enhance weak target detection capability, effectively improve the accuracy and reliability of target recognition, filter out radar noise, improve the signal-to-noise ratio, and realize long-distance vehicle sensing and road detection.

[0019] 2. This invention determines vehicle length based on frame count statistics, tracks and extracts vehicle point clouds in the target area, classifies vehicle types based on vehicle length and point cloud distribution characteristics, and constructs vehicle contour models. This solves the problem of speedometers limiting speeds for different types of vehicles, achieves vehicle length estimation and scattering center localization, has a simple execution structure and lower operating costs, and improves the accuracy of vehicle detection and capture.

[0020] 3. This invention acquires vehicle speed information by accumulating difference-frequency Doppler spectra, converting frequencies and speeds, and performing feature analysis on vehicle trajectory data. This effectively tracks vehicle trajectories within the radar monitoring range, achieving high-precision road trajectory modeling and speed measurement, preventing vehicle speeding, assisting in optimizing vehicle travel paths, and improving road traffic management capabilities. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the structure of a vehicle driving data analysis system based on big data according to the present invention; Figure 2 This is a schematic diagram illustrating the steps of a vehicle driving data analysis method based on big data according to the present invention. Detailed Implementation

[0022] 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.

[0023] Please see Figures 1 to 2 The present invention provides a technical solution: a vehicle driving data analysis system based on big data, comprising: a radar acquisition module, a signal recognition module, a vehicle positioning module, a motion speed measurement module, and a trajectory capture module; The radar acquisition module is used to build a radar acquisition platform in the road, modulate radio frequency signals, change the elevation angle of the radar surface to adjust adjacent detection domains, receive the feedback signals, process them and store the echoes as I-channel and Q-channel spectra, and record the signal frequency and phase information respectively. The radar acquisition module includes: a signal modulation unit and a dual-channel spectrum unit; The signal modulation unit is used to output an intermediate frequency signal through a piezoelectric oscillator using a 24GHz or 77GHz frequency-modulated continuous wave radar. After receiving the echo, it constructs a semantic segmentation network, removes non-road semantic interference, suppresses clutter in the target echo, extracts feature parameters from the Doppler spectral domain, divides the road area, and filters out roadside interference. The dual-channel spectral unit is used to calculate the cross-correlation coefficient of the dominant components of the signal, identify the radar signal, reconstruct and denoise the signal, and obtain two baseband signals, I and Q, through orthogonal demodulation, which represent the amplitude and phase information of the signal, respectively. The I / Q signals are stored in complex form to form a dual-channel output.

[0024] The signal recognition module is used to extract intermediate frequency signal features in the time domain, construct a feature space, amplify the processed signal at the intermediate frequency, perform AD sampling after amplification, perform FFT fast Fourier transform on the I and Q data, convert the discrete time signals of the I and Q channels into frequency domain signals, set signal strength thresholds for large and small vehicles, count the number of frames greater than the threshold, determine the vehicle length based on the frame count results, predict the physical location of the scattering center, and determine the vehicle target area. The signal recognition module includes: a noise suppression unit, a frequency domain analysis unit, and a target localization unit; The noise suppression unit is used to determine the background subspace of each channel through principal component analysis, construct a projection matrix, project the signal onto the orthogonal background subspace, generate a masking matrix to weight the mixed signal components in the high-dimensional space, multiply each component in the feature space by the component weight, and suppress the background signal. The frequency domain analysis unit is used to delay and add frequency domain signals to improve the sliding window energy ratio and the peak height of the detected signal, analyze the frequency domain signal, reconstruct the loss function based on the echo intensity weighting, establish the relationship between frequency domain characteristics and vehicle scattering structure, input the scattering predictor, and measure the position of the scattering center. The target localization unit is used to scan the prior target area of ​​the previous frame, guide the radar to select target boxes with waveform numbers during the scanning process, monitor the vehicle position and adjust the position deviation, encode the target boxes into a high-dimensional space, use the feature distance as the temporal association cost, and the triplet loss encoder learns the geometric similarity of the matching boxes to achieve three-dimensional vehicle target tracking.

[0025] The vehicle positioning module is used to acquire the two-dimensional Doppler spectrum of the radar on the vehicle target area, and to track and extract the point cloud of the vehicle in the target area through CFAR peak detection and DOA direction detection to obtain the vehicle point cloud. Based on the vehicle length and point cloud distribution characteristics, the vehicle type is classified and a vehicle outline model is constructed. The vehicle positioning module includes: a target scanning unit and a contour modeling unit; The target scanning unit is used to extract the signal features of the transmitted signal as tags to optimize the dual-channel radio frequency, constrain the input of the received signal through the interference suppression autoencoder, and calculate the azimuth and elevation angles of the target based on the two-dimensional Doppler spectrum to obtain the three-dimensional spatial coordinates of each detection point, forming a vehicle point cloud. The contour modeling unit is used to aggregate the point clouds of the same vehicle onto the same time axis, cluster and segment the vehicles, calculate the two-dimensional and three-dimensional clustering errors based on the prior vehicle length during clustering, adjust the clustering radius and minimum number of points, classify the vehicles according to the distribution range of the point cloud, and construct the vehicle contour model based on the classification results.

[0026] The motion speed measurement module is used to illuminate the vehicle outline model, measure the vehicle spectrum, interpolate the Doppler frequency and compensate for acceleration, eliminate the difference frequency Doppler spectrum of the signal through phase cancellation, accumulate the difference frequency Doppler spectrum, and obtain the target vehicle speed information. The motion speed measurement module includes: a data compensation unit and a speed measurement unit; The data compensation unit is used to perform FFT transformation on the frequency domain signal at the same distance, refine the spectrum by interpolation, suppress the difference frequency component by double pulse cancellation, improve the signal-to-noise ratio by coherently accumulating pulse trains, and obtain the difference frequency Doppler spectrum. The velocity measurement unit is used to calculate the target radial velocity according to the Doppler frequency corresponding to the peak value of the difference frequency Doppler spectrum, and to determine the real-time velocity of the target vehicle by converting the frequency and velocity.

[0027] The trajectory capture module is used to collect the driving trajectories of each vehicle on the road, construct a vehicle trajectory dataset, build a BP neural network, output real-time safe speed limits, and use a camera to capture the target when the vehicle speed exceeds the speed limit corresponding to the vehicle type.

[0028] The trajectory capture module includes: a trajectory modeling unit, a driving monitoring unit, and a dynamic speed limiting unit; The trajectory modeling unit is used to transform the coordinates of adjacent radar data, determine the vehicle trajectory, calculate the trajectory similarity through motion sequences, filter out scattering and split trajectories, identify the trajectory of the same target, supplement missing trajectory points, perform feature analysis on vehicle trajectory data, extract continuous trajectories of different targets according to time feature patterns, and formulate status labels. The driving monitoring unit is used to construct a trajectory network. The nodes are composed of radar collection points, and the directed edges are composed of clustered GPS trajectories. It includes all driving paths of the vehicle in the city. When the vehicle passes through an adjacent node, the probability of each path is predicted, and the path with the highest probability is selected as the vehicle's spatial trajectory. The dynamic speed limit unit is used to create a training set based on historical traffic flow data, vehicle operation records, accident records, and road markings. The mean squared error is used as the loss function to train a BP neural network. The unit outputs a safe speed limit based on vehicle trajectory, number of vehicles, vehicle type, and vehicle speed as input. When the vehicle speed exceeds the safe speed limit, the unit triggers the camera to capture the image.

[0029] In one specific embodiment, this embodiment selects a test section of an eight-lane urban expressway in a city in China. The section is 2.8km long, with a design speed of 80km / h, and includes one interchange entrance and exit. The daily peak hour traffic flow reaches 4200pcu / h, covering various types of vehicles such as small passenger cars, medium-sized buses, and heavy trucks. It has typical urban expressway traffic flow characteristics and is used to fully verify the operation effect and technical performance of this system.

[0030] In this embodiment, a radar acquisition platform was built at 500m intervals along the test road section, with a total of 6 acquisition nodes deployed along the entire road section. Each platform is equipped with a dual-module system of 24GHz frequency-modulated continuous wave radar and 77GHz millimeter-wave radar, along with an industrial-grade gigabit switch, edge computing terminal, and high-definition capture camera. The radar is installed at a height of 6.2m above the road surface, with the horizontal installation angle perpendicular to the lane direction. The initial pitch angle is set to 12°, which can be dynamically adjusted within the range of 0° to 30° via an electric adjustment mechanism to achieve seamless connection of the detection domains of adjacent acquisition nodes. The effective detection range of a single radar is 5m to 250m, with a distance resolution of 0.3m, an angular resolution of 1.2°, and a maximum speed measurement range of 0 to 180km / h.

[0031] After system startup, the signal modulation unit of the radar acquisition module outputs an intermediate frequency (IF) signal with a center frequency of 24 GHz and a bandwidth of 250 MHz via a piezoelectric oscillator. After power amplification, the signal is radiated to the road surface detection area through the transmitting antenna, while simultaneously receiving echo signals reflected from vehicle targets. To address non-road surface interference such as roadside guardrails, vegetation, and billboards, the system constructs a U-Net-based semantic segmentation network to perform pixel-level semantic segmentation of the range-Doppler spectrum of the echo signal, removing semantic interference from non-road surface areas. Clutter suppression is then achieved through constant false alarm rate (CFAR) detection. Feature parameters such as echo intensity, Doppler frequency shift, and range gate are extracted from the Doppler spectral domain to accurately delineate the effective road detection area, filtering out static roadside interference and sidelobe clutter. Actual measurements show that this step can improve the clutter suppression ratio of the echo signal to over 38 dB, and the static interference filtering rate reaches 96.2%. Simultaneously, the dual-channel spectral unit calculates the cross-correlation coefficient of the dominant components of the preprocessed echo signal, identifies the valid radar echo signal, completes signal reconstruction through wavelet threshold denoising, and then obtains I and Q baseband signals through orthogonal demodulation. The I channel signal records the amplitude information of the echo, and the Q channel signal records the phase information of the echo. Finally, the dual-channel signal is stored and output in complex form I+jQ. In this embodiment, the I / Q dual-channel sampling rate is set to 10MHz, the sampling bit depth is 16bit, the number of sampling points per frame is 1024, the signal-to-noise ratio after signal reconstruction is improved to 26dB, and the phase measurement error is controlled within ±2°.

[0032] After signal acquisition and preprocessing, the noise suppression unit of the signal recognition module decomposes the background echo signal of more than 100 frames through principal component analysis, determines the background subspace of each channel, constructs the corresponding projection matrix, projects the real-time acquired I / Q mixed signal onto the signal subspace orthogonal to the background subspace, generates an adaptive masking matrix to weight the mixed signal components in the high-dimensional space, and multiplies each component in the feature space by the corresponding component weight to achieve effective suppression of the background signal. According to actual measurements, this processing can stabilize the background signal suppression rate at more than 92% and effectively preserve the weak echo signal of the vehicle target. Subsequently, the frequency domain analysis unit amplifies the preprocessed time-domain intermediate frequency signal with a 60dB gain. After amplification, it completes 2-channel synchronous AD sampling and performs 1024-point FFT fast Fourier transform on the discrete-time signals of the I and Q channels respectively, converting the time-domain signals into frequency-domain signals. Then, the frequency-domain signals are processed by delay and addition to improve the sliding window energy ratio and the peak height of the detection signal. In this embodiment, after delay and addition processing, the peak height of the signal is increased by 12.8dB, which effectively improves the detection capability of weak targets. At the same time, a reconstruction loss function is constructed based on echo intensity weighting to establish the mapping relationship between frequency domain features and vehicle scattering structure. This is input into the pre-trained scattering predictor to accurately measure the position of the vehicle scattering center, with the positioning error controlled within 0.4m. The target localization unit scans the prior target region of the previous frame for the detected vehicle target, guides the radar to lock the target box with a unique waveform number during the scanning process, monitors the vehicle position in real time and adjusts the position deviation, encodes the target box into a high-dimensional feature space, uses the feature distance as the temporal association cost, and learns the geometric similarity of the matching box through a triplet loss encoder to achieve continuous tracking of the three-dimensional vehicle target. In this embodiment, the continuous tracking frame rate of the vehicle target is 20fps, the multi-target tracking accuracy MOTA reaches 95.7%, the ID switching rate is less than 0.8%, and it can stably achieve real-time tracking of up to 32 parallel vehicle targets in 8 lanes in the same direction.

[0033] The target scanning unit of the vehicle positioning module extracts core features such as the frequency modulation slope, bandwidth, and center frequency of the transmitted signal as tags, optimizes the parameter matching of the dual-channel radio frequency link, and constrains the input features of the received signal through interference suppression autoencoder to further suppress co-channel interference and multipath reflection interference. For the two-dimensional Doppler spectrum of the effective vehicle target, the azimuth and elevation angles of the target are calculated through CFAR peak detection and DOA direction detection, and the three-dimensional spatial coordinates of each detection point are obtained to form vehicle point cloud data. In this embodiment, a single radar can generate no less than 120 effective point cloud data for a single vehicle, and the point cloud positioning accuracy is ≤0.5m laterally and ≤0.3m longitudinally. The contour modeling unit aggregates point cloud data of the same vehicle onto the same time axis and uses the DBSCAN algorithm to cluster and segment the vehicle point cloud. During the clustering process, based on the prior value of vehicle length output by the signal recognition module, the 2D and 3D clustering errors are calculated, and the clustering radius and minimum point number threshold are dynamically adjusted. For small passenger cars, the clustering radius is set to 1.2m and the minimum point number threshold is 15; for medium-sized buses and heavy trucks, the clustering radius is set to 2.5m and the minimum point number threshold is 30. Based on the distribution range of the point cloud, vehicle type classification is completed, with three categories: small cars, medium-sized cars, and large cars. The classification accuracy reaches 98.3%. At the same time, a 3D contour model of the vehicle is constructed based on the clustered point cloud contour, and the relative error between the model contour size and the actual vehicle size is ≤3.5%.

[0034] The data compensation unit of the motion speed measurement module performs a 2048-point FFT transformation on the frequency domain signal within the same distance gate, refines the spectrum through cubic spline interpolation, and improves the spectral resolution to 500Hz. Then, it suppresses static clutter and low-speed difference frequency components through dual-pulse cancellation, and improves the signal-to-noise ratio of the echo signal by more than 12dB through coherent accumulation of 16 pulse trains, obtaining a high-resolution difference frequency Doppler spectrum. At the same time, for vehicle acceleration and deceleration scenarios, it eliminates the difference frequency Doppler spectrum broadening of the signal through acceleration compensation and phase cancellation. In this embodiment, for moving targets with acceleration ≤5m / s², the Doppler spectrum broadening suppression rate reaches 94.6%, effectively ensuring the speed measurement accuracy. The speed measurement unit calculates the radial velocity of the target based on the Doppler frequency corresponding to the peak value of the difference-frequency Doppler spectrum, and completes the frequency-to-velocity conversion to obtain the real-time driving speed of the target vehicle. In actual measurements, the absolute speed measurement error is ≤±1km / h for vehicles in the speed range of 0~120km / h, and ≤±2km / h for vehicles in the speed range of 120~180km / h. The speed measurement data update frequency is 20Hz, which can meet the real-time monitoring needs of vehicle speed on urban expressways.

[0035] The trajectory modeling unit of the trajectory capture module performs coordinate transformation and spatiotemporal registration on vehicle data from adjacent radar acquisition nodes, completing the trajectory stitching of the same vehicle across the entire road segment. It calculates trajectory similarity using the cosine similarity of motion sequences, filters out false trajectories caused by scattering splits, identifies continuous trajectories of the same target, supplements lost trajectory points using Kalman filtering, performs feature analysis on vehicle trajectory data, extracts continuous trajectories of different targets based on temporal feature patterns, and formulates status labels such as normal driving, speeding, lane changing, and slow driving. In this embodiment, the success rate of continuous vehicle trajectory stitching reaches 99.1%, and the trajectory point completion rate reaches 97.8%, enabling full-journey trajectory tracking of vehicles across the entire road segment. The driving monitoring unit constructs a trajectory network for the entire road segment. The network nodes are composed of the spatial coordinates of radar acquisition points, and the directed edges are composed of the clustered vehicle GPS trajectory and radar trajectory fitting. It covers all feasible driving paths of the vehicle in the test road segment. When the vehicle passes an adjacent node, the probability of the vehicle driving on each path is predicted based on the long short-term memory network. The path with the highest probability is selected as the vehicle's spatial driving trajectory. The average accuracy of trajectory prediction reaches 93.4%, which can predict the vehicle's driving behavior in advance. The dynamic speed limit unit uses a training set based on nearly two years of historical traffic flow data, vehicle operation records, accident records, and road alignment annotation data of the test road section. A three-layer backpropagation (BP) neural network is trained using mean squared error as the loss function. The input layer of the neural network includes four types of feature parameters: vehicle trajectory, real-time traffic flow, vehicle type classification, and real-time vehicle speed. The hidden layer has 64 neurons, and the output layer is the road safety speed limit value under the corresponding operating conditions. The average absolute error of the safety speed limit prediction by this neural network is ≤ ±2 km / h, and the model inference time is ≤ 8 ms, enabling real-time dynamic updates of the safety speed limit. When the system detects that a vehicle's real-time speed exceeds the corresponding vehicle type's safety speed limit by more than 5%, it immediately activates a high-definition camera at the corresponding location to capture panoramic and close-up images of the target vehicle. The captured images have a resolution of 4 megapixels, a shutter response time of ≤ 0.1 s, and an overspeed capture accuracy rate of 98.7%, with no missed captures. Simultaneously, the captured data, vehicle trajectory, speed data, and vehicle type information are uploaded to the traffic control platform, completing the full-process recording of violation data.

[0036] This embodiment completed a 72-hour continuous uninterrupted operation test of the system, monitoring a total of 128,000 vehicles, including 102,000 small vehicles, 11,000 medium vehicles, and 15,000 large vehicles. The overall online rate of the system reached 99.92%, and the vehicle detection rate reached 99.5%. All performance indicators met the engineering application requirements for high-precision analysis of urban road vehicle driving data and traffic control.

[0037] A vehicle driving data analysis method based on big data includes the following steps: Step S1. Build a radar acquisition platform, modulate the radio frequency signal for transmission, receive the echo signal, process it, and store the echo as I-channel and Q-channel spectra, recording the signal frequency and phase information respectively; Step S1 includes: Step S11. Using a 24GHz or 77GHz frequency-modulated continuous wave radar, the intermediate frequency signal is output through a piezoelectric oscillator. After receiving the echo, a semantic segmentation network is constructed to remove non-road semantic interference, suppress clutter in the target echo, extract feature parameters from the Doppler spectral domain, divide the road area, and filter out roadside interference. Step S12 calculates the cross-correlation coefficient of the dominant components of the signal, identifies the radar signal, reconstructs and denoises the signal, and obtains two baseband signals, I and Q, through orthogonal demodulation, which represent the amplitude and phase information of the signal, respectively. The I / Q signals are stored in complex form to form a dual-channel output.

[0038] Step S2. Extract the intermediate frequency signal features, construct the feature space, amplify the processed signal and perform AD sampling, perform FFT transformation on the two data streams to convert the discrete-time signal to the frequency domain signal, set the signal strength threshold, count the number of frames greater than the threshold, determine the vehicle length based on the frame count results, predict the physical location of the scattering center, and determine the target area. Step S2 includes: Step S21. Determine the background subspace of each channel through principal component analysis, construct a projection matrix, project the signal onto the orthogonal background subspace, generate a masking matrix to weight the mixed signal components in the high-dimensional space, multiply each component in the feature space by the component weight, and suppress the background signal. Step S22. Delay and sum the frequency domain signals to increase the sliding window energy ratio and the peak height of the detection signal. Analyze the frequency domain signal, reconstruct the loss function based on the echo intensity weight, establish the relationship between the frequency domain characteristics and the vehicle scattering structure, input the data into the scattering predictor, and measure the position of the scattering center. Step S23. Scan the prior target region of the previous frame, guide the radar to select target boxes with waveform numbers during the scanning process, monitor the vehicle position and adjust the position deviation, encode the target boxes into a high-dimensional space, use the feature distance as the temporal association cost, and use the triplet loss encoder to learn the geometric similarity of the matching boxes to achieve three-dimensional vehicle target tracking.

[0039] In one specific embodiment, a 77GHz frequency-modulated continuous wave radar is used and deployed on a gantry of a six-lane dual carriageway expressway in the city. The installation height is 6m, the radar sampling rate is 10MHz, the number of sampling points per frame is 1024, the frame rate is 20fps, and the effective detection range is 5~200m. This fully verifies the performance of the entire process of step S2.

[0040] In step S21, 120 frames of background echoes from a vehicle-free scene are first acquired. A 4D background subspace is obtained through principal component analysis. An orthogonal projection matrix is ​​constructed, and the real-time I / Q dual-channel signals are projected onto a signal subspace orthogonal to the background. An adaptive masking matrix is ​​then generated to weight the high-dimensional mixed signal components. Actual measurements show that this step achieves a background signal suppression rate of 93%, and the echo signal-to-noise ratio is improved from 12dB to 24dB, effectively eliminating static clutter from roadside guardrails and greenery while fully preserving the weak scattering echo components from vehicles.

[0041] In step S22, the preprocessed time-domain signal is amplified by 60dB and sampled using a 16-bit dual-channel AD converter. It is then converted to a frequency-domain signal via a 1024-point FFT. The frequency-domain signals from four consecutive sliding windows are delayed and summed, resulting in a 13dB increase in peak signal height and a 9-fold increase in sliding window energy ratio. Simultaneously, a reconstruction loss function is constructed based on echo intensity weighting to establish a mapping relationship between frequency-domain features and vehicle scattering structure. This function is input into a pre-trained scattering predictor, and the positioning error of the vehicle scattering center is measured to be ≤0.35m. This accurately distinguishes between strong scattering points at the front and rear of the vehicle. Combined with the statistical results of the number of frames exceeding the threshold, the vehicle length judgment error is ≤±0.4m.

[0042] During step S23, the radar beam scan is guided by the prior target region of the previous frame to lock onto the vehicle target box with a unique waveform number. Vehicle position deviations are corrected in real time, and the target box is encoded into a 128-dimensional high-dimensional feature space. Using feature distance as the temporal association cost, the geometric similarity of the matching boxes is learned through a triplet loss encoder. The measured multi-target tracking accuracy (MOTA) reaches 95.2%, with an ID switching rate of less than 0.7%. It can stably track 24 parallel vehicle targets in six lanes in the same direction, with a 3D target tracking position update delay of ≤50ms and a target region locking accuracy of 99.2%.

[0043] This embodiment has been running stably for 24 hours, processing a total of 32,000 vehicles, and all indicators meet the application requirements for real-time vehicle detection and tracking on urban roads.

[0044] Step S3. Obtain the two-dimensional Doppler spectrum on the target area, and extract the point cloud of the vehicle in the target area by CFAR peak detection and DOA direction detection to obtain the vehicle point cloud. Classify the vehicle type according to the vehicle length and point cloud distribution characteristics, and construct the vehicle outline model. Step S3 includes: Step S31. Extract the signal features of the transmitted signal as tags to optimize the dual-channel radio frequency. Constrain the input of the received signal by the interference suppression autoencoder. Calculate the azimuth and elevation angles of the target based on the two-dimensional Doppler spectrum to obtain the three-dimensional spatial coordinates of each detection point and form a vehicle point cloud. Step S32. Aggregate the point clouds of the same vehicle to the same time axis, cluster and segment the vehicles. During clustering, calculate the two-dimensional and three-dimensional clustering errors based on the prior vehicle length, adjust the clustering radius and minimum number of points, classify the vehicles according to the distribution range of the point cloud, and construct the vehicle outline model based on the classification results.

[0045] Step S4. Illuminate the vehicle contour model with radar, measure the vehicle spectrum, interpolate the Doppler frequency and compensate for acceleration, eliminate the difference frequency Doppler spectrum of the signal through phase cancellation, accumulate the difference frequency Doppler spectrum, and obtain the target vehicle speed information. Step S4 includes: Step S41. Perform FFT transformation on the frequency domain signal at the same distance, refine the spectrum by interpolation, suppress the difference frequency component by double pulse cancellation, improve the signal-to-noise ratio by coherently accumulating pulse trains, and obtain the difference frequency Doppler spectrum; Step S42. Based on the Doppler frequency corresponding to the peak value of the difference frequency Doppler spectrum, calculate the radial velocity of the target vehicle according to the Doppler principle, convert the frequency and velocity, and determine the real-time velocity of the target vehicle.

[0046] Step S5. Collect the driving trajectories of each vehicle on the road, construct a vehicle trajectory dataset, train a neural network, output real-time safe speed limits, and use a camera to capture images of the target when the vehicle speed exceeds the speed limit corresponding to the vehicle type.

[0047] Step S5 includes: Step S51. Transform the coordinates of adjacent radar data to determine the vehicle trajectory, calculate the trajectory similarity through motion sequence, filter out scattering split trajectories, identify the trajectory of the same target, supplement the missing trajectory points, perform feature analysis on the vehicle trajectory data, extract the continuous trajectories of different targets according to the time feature pattern, and formulate status labels. Step S52. Construct a trajectory network. The nodes are composed of radar collection points, and the directed edges are composed of clustered GPS trajectories. It includes all driving paths of the vehicle in the city. When the vehicle passes through an adjacent node, predict the probability of each path and select the path with the highest probability as the vehicle's spatial trajectory. Step S53. Create a training set based on historical traffic flow data, vehicle operation records, accident records, and road markings. Train a BP neural network using mean squared error as the loss function. Analyze vehicle trajectory, number of vehicles, vehicle type, and vehicle speed as inputs to output a safe speed limit. When a vehicle's speed exceeds the safe speed limit, trigger the camera to capture the image.

[0048] In one specific embodiment, a radar sensor speed measurement platform is set up on the road. A piezoelectric oscillator modulates the transmitted signal, and a receiving antenna captures the echo signal. The echo signal is mixed with the transmitted signal to obtain an intermediate frequency signal. The intermediate frequency signal is digitally down-converted and mixed with the quadrature local oscillator signal of a local numerically controlled oscillator. After low-pass filtering, two baseband signals, I and Q, are obtained. The I / Q sampling points within the period are subjected to FFT transformation to obtain the Doppler spectrum signal. Echoes from non-road areas are filtered using a binary mask, static backgrounds are removed using a multi-frame averaging method, target echo clutter is suppressed, amplitude envelopes, phase change rates, and Doppler frequency features of the I / Q signals are extracted, a feature space is constructed, and the received signal can be represented as a linear combination of components. Background noise is removed using a masking matrix, distance cells exceeding a low threshold are counted, vehicle length is estimated based on the number of consecutive distance cells traversed by the vehicle target, the target area is located using a scattering predictor, point clouds are extracted using CFAR detection and DOA estimation, the point clouds are clustered to obtain a vehicle contour model, coherence is accumulated and vehicle speed is measured, a safe speed limit is determined using a neural network, and overspeed capture is triggered.

[0049] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0050] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A vehicle driving data analysis method based on big data, characterized in that, The method includes the following steps: Step S1. Build a radar acquisition platform, modulate the radio frequency signal for transmission, receive the echo signal, process it, and store the echo as I-channel and Q-channel spectra, recording the signal frequency and phase information respectively; Step S2. Extract the intermediate frequency signal features, construct the feature space, amplify the processed signal and perform AD sampling, perform FFT transformation on the two data streams to convert the discrete-time signal to the frequency domain signal, set the signal strength threshold, count the number of frames greater than the threshold, determine the vehicle length based on the frame count results, predict the physical location of the scattering center, and determine the target area. Step S3. Obtain the two-dimensional Doppler spectrum on the target area, and extract the point cloud of the vehicle in the target area by CFAR peak detection and DOA direction detection to obtain the vehicle point cloud. Classify the vehicle type according to the vehicle length and point cloud distribution characteristics, and construct the vehicle outline model. Step S4. Illuminate the vehicle contour model with radar, measure the vehicle spectrum, interpolate the Doppler frequency and compensate for acceleration, eliminate the difference frequency Doppler spectrum of the signal through phase cancellation, accumulate the difference frequency Doppler spectrum, and obtain the target vehicle speed information. Step S5. Collect the driving trajectories of each vehicle on the road, construct a vehicle trajectory dataset, train a neural network, output real-time safe speed limits, and use a camera to capture images of the target when the vehicle speed exceeds the speed limit corresponding to the vehicle type.

2. The vehicle driving data analysis method based on big data according to claim 1, characterized in that: Step S1 includes: Step S11. Using a 24GHz or 77GHz frequency-modulated continuous wave radar, the intermediate frequency signal is output through a piezoelectric oscillator. After receiving the echo, a semantic segmentation network is constructed to remove non-road semantic interference, suppress clutter in the target echo, extract feature parameters from the Doppler spectral domain, divide the road area, and filter out roadside interference. Step S12 calculates the cross-correlation coefficient of the dominant components of the signal, identifies the radar signal, reconstructs and denoises the signal, and obtains two baseband signals, I and Q, through orthogonal demodulation, which represent the amplitude and phase information of the signal, respectively. The I / Q signals are stored in complex form to form a dual-channel output.

3. The vehicle driving data analysis method based on big data according to claim 2, characterized in that: Step S2 includes: Step S21. Determine the background subspace of each channel through principal component analysis, construct a projection matrix, project the signal onto the orthogonal background subspace, generate a masking matrix to weight the mixed signal components in the high-dimensional space, multiply each component in the feature space by the component weight, and suppress the background signal. Step S22. Delay and sum the frequency domain signals to increase the sliding window energy ratio and the peak height of the detection signal. Analyze the frequency domain signal, reconstruct the loss function based on the echo intensity weight, establish the relationship between the frequency domain characteristics and the vehicle scattering structure, input the data into the scattering predictor, and measure the position of the scattering center. Step S23. Scan the prior target region of the previous frame, guide the radar to select target boxes with waveform numbers during the scanning process, monitor the vehicle position and adjust the position deviation, encode the target boxes into a high-dimensional space, use the feature distance as the temporal association cost, and use the triplet loss encoder to learn the geometric similarity of the matching boxes to achieve three-dimensional vehicle target tracking.

4. The vehicle driving data analysis method based on big data according to claim 3, characterized in that: Step S3 includes: Step S31. Extract the signal features of the transmitted signal as tags to optimize the dual-channel radio frequency. Constrain the input of the received signal by the interference suppression autoencoder. Calculate the azimuth and elevation angles of the target based on the two-dimensional Doppler spectrum to obtain the three-dimensional spatial coordinates of each detection point and form a vehicle point cloud. Step S32. Aggregate the point clouds of the same vehicle to the same time axis, cluster and segment the vehicles, calculate the two-dimensional and three-dimensional clustering errors based on the prior vehicle length during clustering, adjust the clustering radius and minimum number of points, classify the vehicles according to the distribution range of the point cloud, and construct the vehicle outline model based on the classification results; Step S4 includes: Step S41. Perform FFT transformation on the frequency domain signal at the same distance, refine the spectrum by interpolation, suppress the difference frequency component by double pulse cancellation, improve the signal-to-noise ratio by coherently accumulating pulse trains, and obtain the difference frequency Doppler spectrum; Step S42. Based on the Doppler frequency corresponding to the peak value of the difference frequency Doppler spectrum, calculate the radial velocity of the target vehicle according to the Doppler principle, convert the frequency and velocity, and determine the real-time velocity of the target vehicle.

5. The vehicle driving data analysis method based on big data according to claim 4, characterized in that: Step S5 includes: Step S51. Transform the coordinates of adjacent radar data to determine the vehicle trajectory, calculate the trajectory similarity through motion sequence, filter out scattering split trajectories, identify the trajectory of the same target, supplement the missing trajectory points, perform feature analysis on the vehicle trajectory data, extract the continuous trajectories of different targets according to the time feature pattern, and formulate status labels. Step S52. Construct a trajectory network. The nodes are composed of radar collection points, and the directed edges are composed of clustered GPS trajectories. It includes all driving paths of the vehicle in the city. When the vehicle passes through an adjacent node, predict the probability of each path and select the path with the highest probability as the vehicle's spatial trajectory. Step S53. Create a training set based on historical traffic flow data, vehicle operation records, accident records, and road markings. Train a BP neural network using mean squared error as the loss function. Analyze vehicle trajectory, number of vehicles, vehicle type, and vehicle speed as inputs to output a safe speed limit. When a vehicle's speed exceeds the safe speed limit, trigger the camera to capture the image.

6. A vehicle driving data analysis system based on big data, characterized in that, The system includes the following modules: Radar acquisition module, signal recognition module, vehicle positioning module, motion speed measurement module, and trajectory capture module; The radar acquisition module is used to build a radar acquisition platform in the road, modulate radio frequency signals, change the elevation angle of the radar surface to adjust adjacent detection domains, receive the feedback signals, process them and store the echoes as I-channel and Q-channel spectra, and record the signal frequency and phase information respectively. The signal recognition module is used to extract intermediate frequency signal features in the time domain, construct a feature space, amplify the processed signal at the intermediate frequency, perform AD sampling after amplification, perform FFT fast Fourier transform on the I and Q data, convert the discrete time signals of the I and Q channels into frequency domain signals, set signal strength thresholds for large and small vehicles, count the number of frames greater than the threshold, determine the vehicle length based on the frame count results, predict the physical location of the scattering center, and determine the vehicle target area. The vehicle positioning module is used to acquire the two-dimensional Doppler spectrum of the radar on the vehicle target area, and to track and extract the point cloud of the vehicle in the target area through CFAR peak detection and DOA direction detection to obtain the vehicle point cloud. Based on the vehicle length and point cloud distribution characteristics, the vehicle type is classified and a vehicle outline model is constructed. The motion speed measurement module is used to illuminate the vehicle outline model, measure the vehicle spectrum, interpolate the Doppler frequency and compensate for acceleration, eliminate the difference frequency Doppler spectrum of the signal through phase cancellation, accumulate the difference frequency Doppler spectrum, and obtain the target vehicle speed information. The trajectory capture module is used to collect the driving trajectories of each vehicle on the road, construct a vehicle trajectory dataset, build a BP neural network, output real-time safe speed limits, and use a camera to capture the target when the vehicle speed exceeds the speed limit corresponding to the vehicle type.

7. A vehicle driving data analysis system based on big data according to claim 6, characterized in that: The radar acquisition module includes: a signal modulation unit and a dual-channel spectrum unit; The signal modulation unit is used to output an intermediate frequency signal through a piezoelectric oscillator using a 24GHz or 77GHz frequency-modulated continuous wave radar. After receiving the echo, it constructs a semantic segmentation network, removes non-road semantic interference, suppresses clutter in the target echo, extracts feature parameters from the Doppler spectral domain, divides the road area, and filters out roadside interference. The dual-channel spectral unit is used to calculate the cross-correlation coefficient of the dominant components of the signal, identify the radar signal, reconstruct and denoise the signal, and obtain two baseband signals, I and Q, through orthogonal demodulation, which represent the amplitude and phase information of the signal, respectively. The I / Q signals are stored in complex form to form a dual-channel output.

8. A vehicle driving data analysis system based on big data according to claim 7, characterized in that: The signal recognition module includes: a noise suppression unit, a frequency domain analysis unit, and a target localization unit; The noise suppression unit is used to determine the background subspace of each channel through principal component analysis, construct a projection matrix, project the signal onto the orthogonal background subspace, generate a masking matrix to weight the mixed signal components in the high-dimensional space, multiply each component in the feature space by the component weight, and suppress the background signal. The frequency domain analysis unit is used to delay and add frequency domain signals to improve the sliding window energy ratio and the peak height of the detected signal, analyze the frequency domain signal, reconstruct the loss function based on the echo intensity weighting, establish the relationship between frequency domain characteristics and vehicle scattering structure, input the scattering predictor, and measure the position of the scattering center. The target localization unit is used to scan the prior target area of ​​the previous frame, guide the radar to select target boxes with waveform numbers during the scanning process, monitor the vehicle position and adjust the position deviation, encode the target boxes into a high-dimensional space, use the feature distance as the temporal association cost, and the triplet loss encoder learns the geometric similarity of the matching boxes to achieve three-dimensional vehicle target tracking.

9. A vehicle driving data analysis system based on big data according to claim 8, characterized in that: The vehicle positioning module includes: a target scanning unit and a contour modeling unit; The target scanning unit is used to extract the signal features of the transmitted signal as tags to optimize the dual-channel radio frequency, constrain the input of the received signal through the interference suppression autoencoder, and calculate the azimuth and elevation angles of the target based on the two-dimensional Doppler spectrum to obtain the three-dimensional spatial coordinates of each detection point, forming a vehicle point cloud. The contour modeling unit is used to aggregate the point clouds of the same vehicle onto the same time axis, cluster and segment the vehicles, calculate the two-dimensional and three-dimensional clustering errors based on the prior vehicle length during clustering, adjust the clustering radius and minimum number of points, classify the vehicles according to the distribution range of the point cloud, and construct the vehicle contour model based on the classification results. The motion speed measurement module includes: a data compensation unit and a speed measurement unit; The data compensation unit is used to perform FFT transformation on the frequency domain signal at the same distance, refine the spectrum by interpolation, suppress the difference frequency component by double pulse cancellation, improve the signal-to-noise ratio by coherently accumulating pulse trains, and obtain the difference frequency Doppler spectrum. The velocity measurement unit is used to calculate the target radial velocity according to the Doppler frequency corresponding to the peak value of the difference frequency Doppler spectrum, and to determine the real-time velocity of the target vehicle by converting the frequency and velocity.

10. A vehicle driving data analysis system based on big data according to claim 9, characterized in that: The trajectory capture module includes: a trajectory modeling unit, a driving monitoring unit, and a dynamic speed limiting unit; The trajectory modeling unit is used to transform the coordinates of adjacent radar data, determine the vehicle trajectory, calculate the trajectory similarity through motion sequences, filter out scattering and split trajectories, identify the trajectory of the same target, supplement missing trajectory points, perform feature analysis on vehicle trajectory data, extract continuous trajectories of different targets according to time feature patterns, and formulate status labels. The driving monitoring unit is used to construct a trajectory network. The nodes are composed of radar collection points, and the directed edges are composed of clustered GPS trajectories. It includes all driving paths of the vehicle in the city. When the vehicle passes through an adjacent node, the probability of each path is predicted, and the path with the highest probability is selected as the vehicle's spatial trajectory. The dynamic speed limit unit is used to create a training set based on historical traffic flow data, vehicle operation records, accident records, and road markings. The mean squared error is used as the loss function to train a BP neural network. The unit outputs a safe speed limit based on vehicle trajectory, number of vehicles, vehicle type, and vehicle speed as input. When the vehicle speed exceeds the safe speed limit, the unit triggers the camera to capture the image.

Images