Unmanned vehicle identification method based on radar perception

Abstract

The invention relates to the technical field of automatic driving and radar signal processing, in particular to an unmanned vehicle identification method based on radar perception. According to the technical scheme of the unmanned vehicle identification method based on radar perception, radar signals emitted by passing vehicles are collected in real time through a receiver installed at an urban traffic intersection or a key position. Secondly, preprocessing and time-frequency analysis are carried out on the collected signals through wavelet transform, and time-domain features and frequency-domain information of the signals are extracted; next, the processed signal image is identified and classified through an improved YOLOv10 target detection algorithm, whether the image is an LFMCW signal of a vehicle-mounted radar is determined, signal information is analyzed, accurate identification and tracking of an unmanned vehicle in urban traffic are realized, and through combination of radar signal processing and a deep learning technology, accurate identification and tracking of the unmanned vehicle in urban traffic are realized. And the limitation of a traditional visual identification method in a complex environment is effectively overcome.

Description

Technical Field

[0001] The present invention relates to the technical fields of autonomous driving and radar signal processing, and particularly relates to a method for identifying driverless vehicles based on radar perception. Background Art

[0002] With the rapid development of autonomous driving technology, the application prospect of driverless vehicles in intelligent transportation systems is gradually becoming broad. However, the identification and tracking of current driverless vehicles still face many challenges. Especially in complex urban traffic environments, traditional vehicle identification methods mostly rely on visual image recognition technology. However, in scenarios such as low light, bad weather, and high-density traffic, the quality and recognition effect of visual images are often severely affected, making it difficult to meet the accuracy requirements in practical applications.

[0003] To overcome the limitations of visual recognition technology, radar perception technology has gradually been applied to the detection and identification of driverless vehicles. The radar system can extract information such as the distance and speed of target objects through reflected waves, and has the advantage of being superior to the visual system under low light and bad weather conditions. Especially FMCW radar, with its high precision and real-time performance, has been widely studied and applied in the fields of traffic monitoring and autonomous driving.

[0004] However, existing radar-based vehicle identification methods often rely on traditional signal processing techniques such as Fourier transform and simple time-domain analysis. These methods have poor robustness in complex environments and are difficult to effectively distinguish driverless vehicles from other traffic targets. At the same time, when extracting features and classifying radar signals, existing methods lack efficient algorithm support and cannot make full use of the advantages of deep learning technology for optimization. Therefore, a method for identifying driverless vehicles based on radar perception can be designed, which can make full use of high-precision receiver technology and, through advanced signal processing and deep learning algorithms, achieve efficient and accurate vehicle identification and tracking in complex urban traffic environments. Summary of the Invention

[0005] To overcome the problem that the existing visual image recognition technology has poor vehicle recognition effect in complex urban traffic environments, especially the defects of low recognition accuracy and poor robustness of traditional visual recognition methods in scenarios such as low light, bad weather, and high-density traffic.

[0006] The technical solution of the present invention is as follows: A method for identifying driverless vehicles based on radar perception, and the specific steps are as follows:

[0007] Step 1: Install a roadside high-precision receiver to collect LFMCW radar signals emitted by passing vehicles in real time;

[0008] Step 2: Preprocess the collected LFMCW signals, perform time-frequency analysis on the signals using wavelet transform, and extract the time-domain features and frequency-domain information of the signals;

[0009] Step 3: Input the preprocessed video images into the improved YOLOv10 deep learning model for training to obtain a network weight model;

[0010] Step 4: Use the network weight model obtained from the training to identify and classify newly collected signal images, and determine whether the images are LFMCW signals emitted by vehicle-mounted radars;

[0011] Step 5: Calculate the time and bandwidth characteristics of the identified LFMCW signals;

[0012] Step 6: Through the identified LFMCW signals and their time and bandwidth characteristics, accurately identify the vehicle successfully.

[0013] Preferably, by collecting LFMCW radar signals emitted by vehicles in real time, performing time-frequency analysis on the signals using wavelet transform, and combining the YOLOv10 deep learning algorithm to perform target detection and classification on the processed time-frequency images. By accurately extracting the time and bandwidth characteristics of the vehicle, the present invention can efficiently identify the target vehicle and achieve accurate identification and tracking of the vehicle. This method can overcome the limitations of traditional visual recognition methods in complex environments, especially performing well under harsh weather conditions such as low light, rain, snow, and haze.

[0014] As a preference, the LFMCW radar signals collected in Step 1 are obtained in real time through a receiver. The waveform of the emitted LFMCW signal is a chirp signal, and its waveform can be expressed as:

[0015]

[0016] where A is the signal amplitude, f 0 is the initial frequency, B is the scanning bandwidth of the chirp, T c is the period of the chirp, and t is the time.

[0017] As a preference, according to the signals in Step 2, time-frequency analysis is performed through wavelet transform to extract time-domain features to support target recognition. According to the characteristics of autonomous driving vehicle-mounted radars, the present invention selects an improved complex Morlet wavelet as the mother wavelet, and its expression is:

[0018]

[0019] where σ is the bandwidth control factor, and its associated relationship with the scale factor a and the desired frequency resolution Δf r satisfies:

[0020]

[0021] where ω 0 is the wavelet center frequency, and Δf r is set according to the radar target recognition requirement to be of the signal bandwidth. By dynamically adjusting the σ parameter, the time-frequency window of the wavelet function is adaptively matched to the local characteristics of the radar signal.

[0022] To meet the radar target separation requirement, it is defined that the time-domain resolution Δt and the frequency-domain resolution Δf of the time-frequency diagram need to satisfy the constraint conditions:

[0023]

[0024] where τ min is the minimum distinguishable target time-delay difference of the radar system, B is the transmission signal bandwidth, and N is an empirical coefficient in the multi-target scenario. Based on the above constraints, the power spectral density P(f) of the input signal s(t) is calculated, and its spectral entropy H = -ΣP(f)log 2 P(f) is used to characterize the signal complexity; subsequently, the objective function is constructed:

[0025] J(a) = αH + β(ΔtΔf)

[0026] where the weight coefficients α = 0.6 and β = 0.4 are determined through Monte Carlo experiments to balance the signal complexity and the resolution requirement. The scale factor a is iteratively adjusted by the gradient descent method to minimize J(a), and at the same time, the time-frequency uncertainty principle is jointly constrained The analytical solution of the optimal scale factor is derived using the Lagrange multiplier method:

[0027]

[0028] In the formula, γ = 0.7 is the time-frequency trade-off factor, T is the signal analysis time window length, and B is the signal bandwidth. This optimal scale factor a opt is directly used for the scale parameter configuration of the subsequent wavelet transform to ensure the theoretical optimality of the time-frequency energy distribution.

[0029] After the parameter optimization is completed, for the size characteristics of the vehicle target and the radar center frequency f 0 , 5, 6 and other levels of details are adaptively selected for multi-scale analysis. The preset scale base values a 5 = 2 5 and a 6 = 2 6 correspond to the theoretical distance resolutions and respectively. To adapt to the actual signal characteristics, through a optDynamically fine-tune the base value and first calculate the correction coefficient (k = 5, 6, is the preset base value), and then set the actual analysis scale to If the corrected a k exceeds the radar distance range [R min , R max (R min and R max are determined by the radar detection ability), then truncation processing is performed according to .

[0030] Perform a convolution operation on the signal through the mother wavelet function ψ(t) to obtain the characteristics of the signal at different scales. Its one-dimensional continuous transform can be expressed as:

[0031]

[0032] where s(t) is the input signal, ψ(t) is the mother wavelet, a k is the actual scale parameter after fine-tuning, and b is the translation factor. Enhance the target saliency through a multi-scale fusion strategy weighted by energy ratio:

[0033]

[0034] Then, perform regularization processing on the fused time-frequency coefficients to suppress noise:

[0035]

[0036] where λ = 0.2 is the noise suppression coefficient, and median b (·) represents taking the median of the squared amplitudes of the wavelet coefficients at all translation positions b under a fixed scale a, and the visibility of weak targets is improved by suppressing background noise. Through this transformation, the time-domain and frequency-domain characteristics of the signal can be effectively extracted, and the time-frequency characteristics of the signal can be analyzed at multiple resolutions.

[0037] Preferably, according to step three, the time-frequency image is input into the YOLOv10 deep learning model for training. The model extracts features and identifies targets from the input image through a convolutional neural network. YOLOv10 uses multi-scale feature maps for target localization and is optimized through the MPDIoU loss function. Its formula is:

[0038]

[0039] where d 1 and d 2 respectively represent the Euclidean distances between the predicted bounding box and the top-left and bottom-right corners of the true bounding box, w and d are the width and height of the true bounding box respectively, and IOU is the traditional intersection over union.

[0040] Preferably, the YOLOv10 model trained in step four is used to classify and identify the newly acquired signal images. For each input image, by comparing each detection result with a preset threshold, it is determined whether the image contains the LFMCW signal emitted by the vehicle-mounted radar.

[0041] Preferably, according to the time-frequency diagram of the identified LFMCW signal in step five, its carrier, time, and bandwidth characteristics are calculated. The bandwidth Δf of the LFMCW signal can be calculated through the frequency modulation rate β and the signal duration T:

[0042] Δf = βT

[0043] The time characteristic can be extracted by calculating the time window of the signal, serving as the basis for subsequent vehicle identification.

[0044] Preferably, according to the time and bandwidth characteristics of the LFMCW signal in step six, the system can efficiently identify the target vehicle and achieve accurate vehicle identification.

[0045] The driverless vehicle identification system includes the above-mentioned driverless vehicle identification method based on radar perception, and also includes a high-precision FMCW radar receiving module, a signal processing unit, a YOLOv10 target detection model, a vehicle identification module, and a data storage and management unit;

[0046] The high-precision FMCW radar receiving module is used to collect the LFMCW radar signals emitted by the passing vehicles in real time;

[0047] The signal processing unit consists of two parts: signal preprocessing and feature extraction parts; the signal preprocessing module is responsible for the preliminary processing of the collected LFMCW signals, including steps such as denoising and wavelet transform, to extract the time-frequency characteristics of the signals and ensure that the details of the signals in the frequency domain and time domain can be fully displayed;

[0048] The feature extraction module then uses wavelet transform to extract the time, bandwidth, and carrier frequency characteristics of the signals, thereby providing valuable information for subsequent target identification;

[0049] The YOLOv10 target detection model, as the core target detection algorithm, is used to perform target identification on the processed video images;

[0050] The YOLOv10 model can classify the LFMCW signal images through deep learning algorithms and accurately determine whether the images are signals emitted by the vehicle-mounted radar;

[0051] The vehicle identification module performs accurate vehicle identification by combining the time and bandwidth characteristics of the LFMCW signal.

[0052] The data storage and management unit is responsible for storing and managing the detection data, constructing a data analysis and processing system. Through the real-time processing of data and the analysis of historical data, the system can continuously optimize the recognition model to ensure efficient operation and high-precision recognition in various complex environments.

[0053] Advantages of the present invention:

[0054] 1. By combining a high-precision receiver and a time-frequency analysis method of wavelet transform, the present invention can effectively extract the time-domain features and frequency-domain information of vehicles, accurately classify and identify the LFMCW signals transmitted by vehicle-mounted radars, ensure accurate vehicle recognition in complex traffic environments, and increase the overall high-precision performance.

[0055] 2. It shows strong robustness in complex environments and can work stably in low-light, bad weather, and high-density traffic scenarios. The improved deep learning algorithm YOLOv10 is used to classify and analyze the signal images, effectively overcoming the limitations of traditional visual recognition technologies in these environments and enhancing strong robustness.

[0056] 3. It can quickly process the signals collected from the receiver and analyze and identify the target vehicle in real time through wavelet transform and deep learning models. The system can maintain efficient recognition and tracking in dynamic traffic scenarios, is applicable to the monitoring and management of driverless vehicles in urban intelligent transportation systems, and improves real-time management performance.

[0057] 4. Compared with traditional high-cost lidar or visual image recognition systems, the present invention is based on radar signal processing and deep learning technologies, has a lower equipment cost, and is easier to install and maintain. The radar equipment has a long service life, can significantly reduce the overall input cost of the intelligent transportation system, and can accurately identify target vehicles under various complex conditions. This technology has wide applicability, is not only applicable to urban traffic management, but also can provide technical support for improving the safety of driverless vehicles. Brief Description of the Drawings

[0058] Figure 1 It is a flowchart of the overall method in the driverless vehicle recognition method of the present invention;

[0059] Figure 2 It is a time-domain diagram of the LFMCW signal after preprocessing in the driverless vehicle recognition method of the present invention;

[0060] Figure 3 It is a time-frequency domain diagram of the LFMCW signal after wavelet transform in the driverless vehicle recognition method of the present invention;

[0061] Figure 4This is the LFMCW signal recognition and detection diagram based on YOLOv10 for the driverless vehicle recognition method of the present invention. Detailed implementation mode

[0062] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0063] The terms used in the present invention are shown in Table 1:

[0064]

[0065] Table 1

[0066] Please refer to Figures 1-4 , the present invention provides an embodiment: a driverless vehicle recognition method based on radar perception, and the specific steps are as follows:

[0067] Step 1: Install a roadside high-precision receiver to collect LFMCW radar signals emitted by passing vehicles in real time;

[0068] Step 2: Preprocess the collected LFMCW signals, perform time-frequency analysis on the signals using wavelet transform, and extract the time-domain features and frequency-domain information of the signals;

[0069] Step 3: Input the preprocessed video image into the improved YOLOv10 deep learning model for training to obtain a network weight model;

[0070] Step 4: Use the network weight model obtained by the training to identify and classify newly collected signal images, and determine whether the images are LFMCW signals emitted by vehicle-mounted radars;

[0071] Step 5: Calculate the time and bandwidth features of the identified LFMCW signals;

[0072] Step 6: Through the identified LFMCW signals and their time and bandwidth features, the accurate recognition of vehicles is successfully achieved.

[0073] When working, a reconnaissance method is adopted. The LFMCW signal transmitted by the on-vehicle radar of the target vehicle is received by the receiver, and signal analysis and classification are carried out on the vehicle. The parameters of its on-vehicle radar are shown in Table 2. The receiving system of the frequency-modulated continuous-wave radar usually includes a receiving antenna, a low-noise amplifier, a mixer, a band-pass filter, and an intermediate-frequency sampling unit. After the received signal is preliminarily amplified by the LNA, it enters the mixer and is mixed with the local oscillator signal to generate an intermediate-frequency signal. These signals extract the effective signals within a specific frequency band through the band-pass filter and are sampled. However, due to the interference of devices in the same frequency band, the received signal may be affected by external interference sources, manifested as phenomena such as elevated background noise and increased false targets. These interference signals may mask the characteristics of the target signal and cause difficulties for subsequent signal analysis.

[0074] Interference signals usually refer to the interference signals emitted by external devices in the same frequency band. These interference signals are superimposed on the target signal after signal mixing, resulting in the system being unable to effectively distinguish the real target from the interference source. In this case, the signal after mixing can be expressed as:

[0075] r(t) = A·cos(2πf 0 t + πβt 2 + φ(t)) + n(t)

[0076] where A is the signal amplitude, f 0 is the initial frequency, β is the frequency modulation rate, t is the time, φ(t) is the phase of the interference signal, and n(t) is the background noise. The superposition of the interference signal will cause the signal-to-noise ratio of the received signal to decrease and may introduce false targets.

[0077] (2) After the received signal passes through the LNA, mixer, and band-pass filter, the system will obtain the signal within the band-pass frequency range. At this time, all interference components below and above the BPF bandwidth will be filtered out, and the signal will only be retained within the band-pass frequency band. However, the signal after mixing may still contain some noise components, especially due to the interference of external co-frequency devices, which may cause the elevation of the background noise. To effectively suppress these noises and extract the target information, the system needs to perform signal preprocessing, including noise removal and filtering.

[0078] (a) Band-pass filtering

[0079] The band-pass filter is used to remove the components in the signal that are not within the set frequency range and retain the signal within the target frequency range. After BPF filtering, the target signal will be concentrated within the intermediate-frequency range, and the interference signal will be eliminated as much as possible. The frequency response of the BPF can be expressed as:

[0080]

[0081] where f cis the center frequency and Δf is the bandwidth.

[0082] (b) Noise removal

[0083] Since the signal may still contain noise components introduced by interference, the Kalman filtering technique is used to denoise the signal. Kalman filtering is a recursive filtering algorithm that can effectively estimate the system state and reduce noise interference. It optimizes the estimation result based on the prediction of the current state and the sensor observation data. Assuming the signal is modeled as a dynamic system, the state equation of the Kalman filter is as follows:

[0084] x k = Ax k-1 + Bu k + w k

[0085] where x k is the state vector of the system, A is the state transition matrix, B is the control matrix, u k is the control input, and w k is the process noise.

[0086] The observation equation is:

[0087] z k = Hx k + v k

[0088] where z k is the observation value, H is the observation matrix, and v k is the observation noise.

[0089] Through the calculation of the Kalman gain K k , the filter can update the state estimation value at each moment and minimize the prediction error:

[0090]

[0091] K k = P k-1 H T (HP k-1 H T + R) -1

[0092] where P k-1 is the error covariance at the previous moment, and R is the covariance matrix of the observation noise.

[0093] The Kalman filter estimates the signal at each moment in a recursive manner, thus effectively removing the background noise and retaining the target signal.

[0094] (3) After the signal preprocessing is completed, the system uses wavelet transform for time-frequency analysis. The purpose is to extract the time-domain features of the signal and provide necessary inputs for target recognition. Compared with the traditional Fourier transform, wavelet transform has good time-frequency localization ability and can analyze the time variation and frequency characteristics of the signal simultaneously. Especially for LFMCW signals, the linear change of its frequency has significant time locality, while the Fourier transform has the problem of frequency ambiguity when dealing with non-stationary signals. Through wavelet transform, the instantaneous features and dynamic changes of the target signal can be revealed more accurately.

[0095] Regarding the characteristics of LFMCW signals emitted by automotive radars for autonomous driving, the present invention selects an improved complex Morlet wavelet as the mother wavelet and utilizes its good time-frequency localization ability to meet the multi-resolution analysis requirements of linear frequency modulation signals. Its expression is:

[0096]

[0097] where σ is the bandwidth control factor, and its correlation relationship with the scale factor a and the expected frequency resolution Δf r satisfies:

[0098]

[0099] where ω is the wavelet center frequency, that is, the radar operating frequency of 77 GHz, and Δf r is set to of the signal bandwidth according to the radar target recognition requirements. By dynamically adjusting the σ parameter, the time-frequency window of the wavelet function can adaptively match the local features of the radar signal.

[0100] To meet the radar target separation requirements, it is defined that the time-domain resolution Δt and the frequency-domain resolution Δf of the time-frequency diagram need to satisfy the constraint conditions:

[0101]

[0102] where is the minimum distinguishable target time delay difference of the radar system, B = 150 MHz is the transmit signal bandwidth, and N = 3 is the empirical coefficient in the multi-target scenario. Based on the above constraints, the power spectral density P(f) of the input signal s(t) is calculated, and its spectral entropy H = -ΣP(f)log 2 P(f) is used to characterize the signal complexity. In the multi-target scenario, a high entropy value (H>3) indicates that the signal energy is dispersed in multiple frequency bands, such as multi-targets or noise interference, and the resolution needs to be improved to separate the targets; a low entropy value (H<2) indicates that the signal is concentrated, and the resolution can be appropriately relaxed to reduce the calculation amount; then the objective function is constructed:

[0103] J(a) = αH + β(ΔtΔf)

[0104] Among them, the weight coefficients α = 0.6 and β = 0.4 are determined through Monte Carlo experiments. That is, on 100 groups of radar echo data sets containing 3 - 5 targets, with the target separation success rate and calculation time consumption as evaluation indicators, it is verified that this weight ratio reaches the optimal balance when the signal - to - noise ratio is greater than 10 dB, which is used to balance the signal complexity and resolution requirements. The scale factor a is iteratively adjusted by the gradient descent method to minimize J(a), while jointly constraining by the time - frequency uncertainty principle The analytical solution of the optimal scale factor is derived using the Lagrange multiplier method:

[0105]

[0106] In the formula, γ = 0.7 is the time - frequency trade - off factor, T = 7.33 μs is the signal analysis time window length, and B = 150 MHz is the signal bandwidth. This optimal scale factor a opt is directly used for the scale parameter configuration of the subsequent wavelet transform to ensure the theoretical optimality of the time - frequency energy distribution.

[0107] After the parameter optimization is completed, for the maximum radar detection distance of 200 m and the range resolution of 1 m, details at levels 5, 6, etc. are adaptively selected for multi - scale analysis. The preset scale base values a 5 = 2 5 and a 6 = 2 6 correspond to the theoretical range resolutions and To adapt to the actual signal characteristics, the base values are dynamically fine - tuned through a opt First, the correction coefficients are calculated to obtain η 5 = 0.004 and η 6 = 0.002. However, because η k is too small and exceeds the physical range, finally a 5 = 2 5 and a 6 = 2 6 are taken to retain the preset resolution characteristics.

[0108] The signal is convolved with the mother wavelet function ψ(t) to obtain the characteristics of the signal at different scales. Its one - dimensional continuous transform can be expressed as:

[0109]

[0110] Among them, s(t) is the input signal, ψ(t) is the mother wavelet, a k is the actual scale parameter after fine - tuning, and b is the translation factor. The target saliency is enhanced through a multi - scale fusion strategy weighted by energy proportion:

[0111]

[0112] Then, the fused time-frequency coefficients are regularized to suppress noise:

[0113]

[0114] where λ = 0.2 is the noise suppression coefficient, and median b (·) represents taking the median of the squared magnitudes of the wavelet coefficients at all translation positions b for a fixed scale a, and the visibility of weak targets is enhanced by suppressing background noise. Through this transformation, the time-domain and frequency-domain characteristics of the signal can be effectively extracted, and the time-frequency characteristics of the signal can be analyzed at multiple resolutions.

[0115] As Figure 3 shown, after completing the time-frequency analysis of the wavelet transform, the system obtains the characteristic information of the signal at different time scales, and this information contains the time-domain characteristics related to the target object, such as the appearance time, disappearance time of the target, and its change pattern. By adaptively selecting this information, the system can provide accurate signal characteristics for subsequent target recognition and classification, thereby improving the accuracy and robustness of target detection.

[0116] (4) After the system adaptively selects the time-frequency features of the signal, it inputs this information into the YOLOv10 deep learning model for training. The time-frequency diagram of the LFMCW signal has significant linear frequency modulation characteristics, and its performance in the time-frequency plane is a set of linear trajectories with obvious rules. However, due to the complex background and strong interference noise often accompanied in the actual application scenario of the signal, this linear feature may be weakened or masked. Therefore, the YOLOv10 model is selected to process such signals, and by accurately regressing the bounding boxes and deeply extracting multi-dimensional features, the subtle differences in the time-frequency diagram are captured, thereby improving the accuracy of signal recognition.

[0117] During the training process of YOLOv10, according to the characteristics of the vehicle-mounted LFMCW signal, the present invention optimizes the loss function of the model, especially introducing an improved method based on MPDIoU in the aspect of bounding box regression. The MPDIoU loss function not only considers the traditional intersection over union, but also introduces the differences in the polar coordinate distances of the corner points between the predicted box and the ground truth box and the box size information. This improvement can significantly improve the accuracy of bounding box regression, enabling the model to more accurately capture the linear trajectory characteristics of the LFMCW signal when identifying it, and still showing good robustness even in complex backgrounds and strong noise.

[0118] Specifically, the calculation formula of MPDIoU is as follows:

[0119]

[0120] Among them, d 1 and d 2 respectively represent the Euclidean distances between the predicted bounding box and the upper - left and lower - right corners of the ground - truth bounding box. w and d are the width and height of the ground - truth bounding box respectively, and IOU is the traditional intersection - over - union. Based on this formula, the bounding - box regression loss term is defined as:

[0121] L MPDIoU = 1 - MPDIoU

[0122] In addition, the system also combines the classification loss L cls and the confidence loss L conf to construct a complete object - detection loss function. The definition of the classification loss L cls is as follows:

[0123]

[0124] Among them, y i represents the true class label of the object, and p i represents the probability of the predicted class. The definition of the confidence loss L conf is:

[0125]

[0126] Among them, t i is the true confidence of the object's existence, and c i is the predicted confidence.

[0127] Finally, the total loss function of the model is:

[0128] L total = L cls + L conf + L MPDIoU

[0129] By introducing MPDIoU and combining classification and confidence losses, the improved loss function significantly enhances the performance of the model in object - detection tasks, especially in the case of complex backgrounds and strong noise signals, showing higher robustness and generalization ability.

[0130] To further improve the model's recognition ability for radar - signal images, the present invention introduces an attention mechanism, especially the spatial - attention mechanism and the channel - attention mechanism. The spatial - attention mechanism can assign higher weights to important regions in the image, enabling the model to focus on the significant regions of LFMCW signals. Specifically, the spatial - attention mechanism processes the input image using average - pooling and max - pooling operations and generates a weight map through convolution:

[0131] S = σ(Conv 3×3(AvgPool(X)+MaxPool(X))

[0132] where σ is the Sigmoid activation function, generating the spatial attention weight map S, Conv 3×3 represents the convolution operation, AvgPool(X) and MaxPool(X) represent the average pooling and max pooling operations respectively, and X is the input feature map. Through this operation, the model can automatically focus on the key signal regions in the radar image, thereby improving the accuracy of target recognition and reducing the interference of background noise.

[0133] During the training process, the present invention optimizes the training parameters to adapt to the characteristics of LFMCW signals. The learning rate adopts an adaptive adjustment strategy, and the initial learning rate is set to η 0 = 1×10 -4 , and is dynamically adjusted through learning rate decay as the training process progresses. The Batch Size is set to 16 to balance the training speed and memory consumption. The Adam optimizer is selected as the optimizer, and its update formula is:

[0134]

[0135] where η is the learning rate, m t is the first moment of the gradient, v t is the second moment of the gradient, and m t is a constant to prevent division by zero. To ensure that the network can be fully trained at the appropriate time, Epoch is set to 300. To avoid overfitting, an early stopping strategy is also introduced during the training process. When the validation set loss fails to effectively decrease within several rounds, the training will automatically stop. In addition, data augmentation methods such as random rotation, scaling, and cropping are also adopted during the training process to enhance the generalization ability of the model.

[0136] (5) After the training is completed, the YOLOv10 model will perform target recognition and classification on new radar signal images. Each input image will undergo feature extraction through the convolutional layer of the model and output the location and category information of the target. During this process, the system filters out the detection results with low confidence by setting a confidence threshold to ensure the accuracy of the detection results. In this patent, the confidence threshold is set to 0.7, which means that only when the prediction confidence of the model exceeds 70% will the target be considered a valid detection result. Through this setting, the situation of misrecognition can be effectively reduced, ensuring that only highly reliable detection results are output.

[0137] To further improve the accuracy of object detection, the system also adopts the non-maximum suppression algorithm to eliminate redundant overlapping bounding boxes. During this process, the system sets an overlapping threshold, which is set to 0.4. That is, when the overlap degree of two candidate bounding boxes exceeds 40%, the system will select the bounding box with a higher confidence as the final detection result and eliminate other bounding boxes with a higher overlap degree. In this way, the detection result can be further optimized to avoid the same object being repeatedly detected by multiple bounding boxes.

[0138] As Figure 4 shown, the system successfully identifies the objects in the radar signal image and accurately calibrates the positions of the objects through bounding boxes. At the same time, the system also outputs the category information of the objects, clearly identifying that the object is the LFMCW signal emitted by an autonomous vehicle. These information will be used as the input for subsequent processing to help achieve vehicle tracking and identification, providing key support for the autonomous driving system.

[0139] By setting appropriate confidence thresholds and IoU thresholds, the method of this patent effectively improves the accuracy of object recognition and classification, ensures the reliability of the system in complex traffic environments, and further enhances the practicality and feasibility of autonomous vehicle recognition.

[0140] (6) After object recognition is completed, the system analyzes the time-frequency characteristics of the LFMCW signal and calculates its time and bandwidth characteristics. These characteristics are crucial for the subsequent accurate identification of vehicles. The bandwidth Δf of the LFMCW signal can be calculated through the frequency modulation rate Δf and the duration β of the signal:

[0141] Δf = βT

[0142] In addition, the system can also extract time characteristics through the time window characteristics of the signal to further judge the dynamic behavior of the object. These time and bandwidth characteristics will help the system better distinguish different types of vehicles and improve the recognition accuracy.

[0143] (7) Combining the time-frequency characteristics, time and bandwidth characteristics extracted in the previous steps, the system can achieve accurate identification of the target vehicle. By comparing the characteristics of the target with the preset vehicle templates, the system can accurately judge relevant information such as the type of the target vehicle. This recognition process is not affected by environmental factors such as weather and lighting, and can perform vehicle recognition stably and efficiently, thus providing accurate decision-making support for subsequent autonomous driving control.

[0144] Through this series of processes, the method for identifying autonomous vehicles based on radar perception of the present invention realizes high-precision and real-time vehicle recognition, can be widely applied in fields such as intelligent transportation and autonomous driving, and provides strong support for the realization of autonomous driving technology.

[0145]

[0146]

[0147] Table 2

[0148] The core of the present invention lies in real-time collecting the LFMCW radar signals emitted by vehicles, performing time-frequency analysis on the signals by using wavelet transform, and combining with the YOLOv10 deep learning algorithm to conduct target detection and classification on the processed time-frequency images. By accurately extracting the time and bandwidth characteristics of vehicles, the present invention can efficiently identify target vehicles and achieve precise identification and tracking of vehicles. This method can overcome the limitations of traditional vision recognition methods in complex environments, especially performing excellently under adverse weather conditions such as low light, rain, snow, and haze.

[0149] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the purpose of the present invention.

Claims

1. A method for identifying unmanned vehicles based on radar perception, characterized in that: The specific steps are as follows: Step 1: Install a high-precision roadside receiver to collect LFMCW radar signals emitted by passing vehicles in real time; Step 2: Preprocess the collected LFMCW signal, perform time-frequency analysis on the signal using wavelet transform, and extract the time domain characteristics and frequency domain information of the signal; Step 3: Input the preprocessed video image into the improved YOLOv10 deep learning model for training to obtain the network weight model; Step 4: using the trained network weight model to identify and classify the newly acquired signal image to determine whether the image is a LFMCW signal emitted by the vehicle-mounted radar; Step 5: Calculate the time and bandwidth characteristics of the identified LFMCW signal; Step 6: The vehicle is successfully identified by identifying the LFMCW signal and its time and bandwidth characteristics.

2. The method for identifying an unmanned vehicle based on radar perception according to claim 1, characterized in that: According to the LFMCW radar signal collected in step 1, the receiver is used to obtain the signal in real time. The waveform of the transmitted LFMCW signal is a chirp signal, and its waveform can be expressed as: Where A is the signal amplitude, f0 is the initial frequency, B is the chirp scanning bandwidth, T c is the period of chirp, and t is the time.

3. The method for identifying an unmanned vehicle based on radar perception according to claim 1, characterized in that: According to the signal in step 2, time-frequency analysis is performed through wavelet transform to extract time domain features to support target recognition. According to the characteristics of the autonomous driving vehicle-mounted radar, the present invention selects the improved complex Morlet wavelet as the mother wavelet, and its expression is: Where σ is the bandwidth control factor, which is related to the scale factor a and the desired frequency resolution Δf r The association relationship satisfies: Among them, ω0 is the center frequency of the wavelet, Δf r Set as signal bandwidth according to radar target recognition requirements By dynamically adjusting the σ parameter, the time-frequency window of the wavelet function is adaptively matched to the local characteristics of the radar signal. In order to meet the radar target separation requirements, the time domain resolution Δt and frequency domain resolution Δf of the time-frequency diagram must meet the constraints: Among them, τ min is the minimum distinguishable target delay difference of the radar system, B is the transmission signal bandwidth, and N is the empirical coefficient in the multi-target scenario. Based on the above constraints, the power spectral density P(f) of the input signal s(t) is calculated, and its spectral entropy H = -∑P(f)log2P(f) is calculated to characterize the signal complexity; then the objective function is constructed: J(a)=αH+β(ΔtΔf) Among them, the weight coefficients α = 0.6 and β = 0.4 are verified and determined by Monte Carlo experiments to balance the signal complexity and resolution requirements. The scale factor a is iteratively adjusted by the gradient descent method to minimize J(a), and the time-frequency uncertainty principle is combined to constrain The optimal scaling factor analytical solution is derived using the Lagrange multiplier method: In the formula, γ = 0.7 is the time-frequency trade-off factor, T is the signal analysis time window length, and B is the signal bandwidth. opt It is directly used for the scale parameter configuration of the subsequent wavelet transform to ensure the theoretical optimality of the time-frequency energy distribution. After the parameter optimization is completed, according to the size characteristics of the vehicle target and the radar center frequency f0, the details of levels 5 and 6 are adaptively selected for multi-scale analysis. The preset scale base value a5 = 2 5 and a6=2 6 Corresponding to the theoretical distance resolution and To adapt to the actual signal characteristics, a opt Dynamically fine-tune the base value and calculate the correction coefficient first is the preset base value), and then the actual analysis scale is set to If the correction a k Out of radar range R min ,R max ](R min , R max Determined by the radar detection capability), then Perform truncation processing. The signal is convolved with the mother wavelet function (t) to obtain the characteristics of the signal at different scales. Its one-dimensional continuous transformation can be expressed as: Among them, s(t) is the input signal, ψ(t) is the mother wavelet, a k is the actual scale parameter after fine-tuning, and b is the translation factor. The multi-scale fusion strategy weighted by energy proportion is used to enhance the target saliency: Then, the fused time-frequency coefficients are regularized to suppress noise: Among them, λ = 0.2 is the noise suppression coefficient, median b (·) represents the median of the square of the wavelet coefficient amplitudes of all translation positions b under a fixed scale a, which improves the visibility of weak targets by suppressing background noise. Through this transformation, the time domain and frequency domain features of the signal can be effectively extracted, and the time-frequency characteristics of the signal can be analyzed at multiple resolutions.

4. The method for identifying an unmanned vehicle based on radar perception according to claim 1, characterized in that: According to step 3, the video image is input into the YOLOv10 deep learning model for training. The model extracts features and recognizes targets from the input image through a convolutional neural network. YOLOv10 uses multi-scale feature maps for target positioning and optimizes it through the MPDIoU loss function, whose formula is: Among them, d1 and d2 represent the Euclidean distance between the upper left corner and the lower right corner of the predicted bounding box and the true bounding box, w and d are the width and height of the true bounding box, and IOU is the traditional intersection-union ratio.

5. The method for identifying an unmanned vehicle based on radar perception according to claim 1, characterized in that: The newly acquired signal images are classified and recognized using the YOLOv10 model trained in step 4. For each input image, each detection result is compared with the preset threshold to determine whether the image contains the LFMCW signal emitted by the vehicle-mounted radar.

6. The method for identifying an unmanned vehicle based on radar perception according to claim 1, characterized in that: According to the time-frequency diagram of the LFMCW signal identified by analysis in step 5, its carrier, time and bandwidth characteristics are calculated. The bandwidth Δf of the LFMCW signal can be calculated by the frequency modulation rate β and the signal duration T: Δf=βT The temporal features can be extracted by calculating the time window of the signal, which serves as the basis for subsequent vehicle identification.

7. The method for identifying an unmanned vehicle based on radar perception according to claim 1, characterized in that: According to the time and bandwidth characteristics of the LFMCW signal combined in step six, the system can efficiently identify the target vehicle and achieve accurate vehicle identification.

8. Unmanned vehicle identification system, characterized by It includes an unmanned vehicle identification method based on radar perception according to claims 1-7, which also includes a high-precision FMCW radar receiving module, a signal processing unit, a YOLOv10 target detection model, a vehicle identification module, and a data storage and management unit; The high-precision FMCW radar receiving module is used to collect the LFMCW radar signals emitted by passing vehicles in real time; The signal processing unit consists of two parts: signal preprocessing and feature extraction; The signal preprocessing module is responsible for the preliminary processing of the collected LFMCW signal, including steps such as denoising and wavelet transform, to extract the time-frequency characteristics of the signal and ensure that the details of the signal in the frequency domain and time domain can be fully displayed; The feature extraction module uses wavelet transform to extract the time, bandwidth and carrier frequency characteristics of the signal, thereby providing valuable information for subsequent target recognition; The YOLOv10 target detection model is used as the core target detection algorithm to identify targets in processed video images; The YOLOv10 model uses a deep learning algorithm to classify LFMCW signal images and accurately determine whether the image is a signal emitted by a vehicle-mounted radar; The vehicle identification module accurately identifies the vehicle by combining the time and bandwidth characteristics of the LFMCW signal; The data storage and management unit is responsible for storing and managing the detection data and building a data analysis and processing system. Through real-time processing of data and analysis of historical data, the system can continuously optimize the recognition model to ensure efficient operation and high-precision recognition in various complex environments.

Images