An automatic parking obstacle perception method based on millimeter-wave radar

Abstract

The present invention belongs to the technical field of automatic parking for automobiles, and particularly relates to an automatic parking obstacle perception method based on millimeter-wave radar. Through data collection, network model construction, and model training, a denoising network and a classification network are ultimately obtained. The denoising network can output denoised target data based on real-time input RDMAP and corresponding operating temperature. After CFAR, speed measurement, angle measurement, and distance interpolation, the position of the target relative to the radar can be calculated. By accumulating multiple frames of RDMAP, the classification network can output the target category. The denoising network eliminates noise energy and improves target distance accuracy. Furthermore, during data set creation and real-time processing, the temperature of the radar RF chip is detected and used as one of the network model inputs, thereby improving the robustness of the denoising network at different temperatures and avoiding temperature-induced noise residue.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic parking technology for automobiles, and in particular to an automatic parking obstacle perception method based on millimeter-wave radar. Background Art

[0002] To avoid accidents, the intelligent driving system must constantly monitor the vehicle's surroundings during automated parking. Millimeter-wave radar, as one of the sensing organs of the intelligent driving system, must possess the ability to accurately perceive the environment. Currently, millimeter-wave radar applications are primarily focused on driving, while ultrasonic radar is often used for environmental perception during parking. Automotive millimeter-wave radar utilizes a frequency-modulated continuous wave (FMCW) system with a zero sampling starting distance. Theoretically, this results in a minimal blind spot, making it useful for close-range target detection. However, in practice, DC signal noise impacts both close-range target detection and ranging performance. This significantly reduces close-range target detection capabilities, balancing false alarm rates with missed detection rates.

[0003] The patent "A Method and System for Automatic Parking Based on 4D Millimeter-Wave Radar" mentions using 4D millimeter-wave radar to detect the reference distance between the rear of the vehicle and various reference objects, but fails to consider the actual close-range target detection performance and ranging accuracy of millimeter-wave radar. The paper "A Parking System Obstacle Detection Method" and the patent "A Method, System, and Device for Automatic Parking Obstacle Distance Detection" both use multiple ultrasonic sensors for obstacle ranging, which offers high ranging accuracy. However, ultrasonic radar only outputs distance information, not target position, altitude, or speed. The patent "A Vehicle-Mounted Millimeter-Wave Radar Target Classification Method, Device, System, and Storage Medium" uses Doppler frequency-time energy distribution maps as network input for target classification. However, during signal preprocessing, vector mean cancellation is used to filter out Doppler-dimensional stationary clutter, which can miss important targets when the radar is stationary or traveling at low speeds. This is particularly true in parking mode, where targets are mostly stationary or moving slowly relative to the radar. Furthermore, this patent uses single-frame images for classification, resulting in low accuracy. The paper "Human Detection Based on Time-Varying Signature on Range-Doppler Diagram Using Deep Neural Networks" uses multiple frames of images as input, which improves classification accuracy. However, it is computationally intensive and consumes a lot of memory, making it unsuitable for in-vehicle equipment. Summary of the Invention

[0004] This invention mainly proposes an automatic parking obstacle perception method based on millimeter-wave radar for automatic parking scenarios, which meets the needs of the intelligent driving system to accurately perceive obstacles around the vehicle body during parking, thereby realizing efficient and safe automatic parking.

[0005] The vehicle-mounted millimeter-wave radar transmits a linear frequency modulated continuous wave signal:

[0006]

[0007] in is the fast time, T is the duration (i.e. pulse width) of a single chirp (which can be understood as a pulse), is the carrier frequency, is the fast time frequency, and correspond, is the modulation frequency. rect represents the rectangular function, and the expression is , exp is the exponential function, the expression is , j represents the imaginary part.

[0008] The target reflected echo is:

[0009]

[0010] in is the distance from the target to the radar array, and c is the speed of light. After receiving the echo, the frequency modulation is processed, that is, it is conjugated and multiplied with the reference signal to obtain a single-frequency signal, and there is , that is, the frequency of the single-frequency signal is proportional to the distance to the target. The echo is processed by deinterlacing and Fourier transforming to obtain a one-dimensional range image, where each sampling frequency corresponds to a target at a different distance.

[0011] The vehicle-mounted radar has a short operating range, and the target echo delay at the maximum operating range is also less than the pulse width. Therefore, the sampling starting distance is zero, and the corresponding starting frequency is also zero.

[0012] The sampling start frequency is zero, which inevitably introduces DC and low-frequency noise into the echo. When the noise power is too high, it can drown out the radar's ability to detect close-range targets, creating a detection blind spot. This blind spot varies with the radar's range resolution; the lower the resolution, the larger the blind spot. In automated parking scenarios, precise perception of the vehicle's surroundings is essential to avoid collisions.

[0013] To solve the problem of radar close-range detection blind spots, the present invention proposes an automatic parking obstacle perception method based on millimeter-wave radar. The specific steps are as follows:

[0014] Step S1: Untargeted dataset creation,

[0015] The system collects 2DFFT results (2DFFT refers to the two-dimensional fast Fourier transform, a core algorithm for processing linear frequency modulated continuous wave (LFM) signals, used to extract target range and velocity information from raw data) for each radar receiving channel in various target-free scenarios (including various open spaces with the radar stationary, various open spaces with the radar moving at low speed, and high-speed moving scenarios (where the radar moves at speeds below 2 m / s and above 10 m / s). The system then extracts data from multiple range sampling units in the radar's close-range region (0-5 meters from the radar, which can be adjusted based on actual applications) for each receiving channel. While collecting data, the radar's RF chip temperature is recorded and retrieved to ensure that the target-free dataset covers as many scenarios as possible and includes all possible operating temperatures (upper and lower limits) of the radar RF chip. Ultimately, this extracted data forms the target-free data sample.

[0016] It should be noted that the distance dimension refers to the dimension in the direction of radar wave propagation, used to measure the straight-line distance between the target and the radar; the unit used is the actual distance represented by each sampling point. The sampling starting distance is the radar array plane, and spatial sampling is performed according to the resolution. The operating temperature of the radar RF chip, not the ambient temperature, is determined by the operating temperature. The radar generates heat during operation, causing the internal temperature to rise. This is affected to a certain extent by the ambient temperature, which is mainly determined by the radar power and heat dissipation performance. The ambient temperature should be considered during radar design, so the operating temperature of the radar RF chip mentioned above takes into account all operating temperatures in extreme environmental conditions.

[0017] Step S2: Target dataset creation,

[0018] Various targets are set up in the radar's close-range area and categorized into strong stationary targets, weak stationary targets, and moving targets based on their characteristics. Moving targets are defined as targets whose relative speed exceeds one radar speed resolution unit. The counterpart to moving targets is stationary targets, which are defined as targets whose relative speed does not exceed one radar speed resolution unit. Stationary targets are further categorized into strong and weak targets, primarily based on the intensity of their electromagnetic wave scattering. Strong stationary targets are defined as strong scattering stationary targets, including but not limited to metal and walls; weak stationary targets are defined as weak scattering stationary targets, including but not limited to cones and green belts.

[0019] Multiple distance angles are set for each target, and the 2DFFT results of the radar's receiving channel distances are collected under the radar motion state of the same target (radar motion states are divided into stationary state and low-speed motion state, and the radar speed during low-speed motion is less than 2m / s). The data of multiple distance dimension sampling units in the radar close-range area of ​​each receiving channel are extracted. While collecting data, the temperature of the radar RF chip is recorded and extracted. The extracted data forms the target data sample, and the set target is labeled with the label corresponding to the target category.

[0020] Step S3: Construct a denoising network, use the target-free data samples from step S1 as training samples, input them into the denoising network, perform training, and output the residual. Continue training iterations until the residual is equal to zero, thereby obtaining a denoising network with a residual equal to zero.

[0021] To prevent DC and low-frequency noise from affecting radar close-range target detection, that is, to eliminate DC and low-frequency noise in the echo, in step S1, data from various target-free scenarios (target-free scenarios include various open scenes with the radar stationary, various open scenes with the radar moving at low speed, and scenes with the radar moving at high speed) are collected. Since there are no targets in the radar close-range area in open scenes and scenes with the radar moving at high speed (the radar moving at high speed is used to collect noise samples under high-speed motion for better denoising), the 2DFFT data collected at this time is DC and low-frequency noise. This noise changes with temperature, so the radar RF chip temperature is also input. Through training, when the above data is input, the network output residual is close to zero (in theory, the residual is equal to zero). This is the purpose and expected result of training, which allows the denoising network to remember the characteristics of the noise.

[0022] Step S4: The target data sample of step S2 is used as a training sample and input into the denoising network finally obtained in step S3 for training. The residual is output and the target detection is performed on the residual result to determine whether the target exists. The target detection position and target category are obtained. The training iteration is continued until the target detection position and target category correspond one-to-one with the target position and target category set in step S2, thereby obtaining the target detection network (i.e., the trained denoising network) and classification network.

[0023] The 2DFFT data collected in step S2 is DC and low-frequency noise + target data. Through training, the denoising network can remove the noise data in this type of data according to the noise characteristics and retain the target data, thereby obtaining the target detection network (i.e., the trained denoising network) and classification network.

[0024] In summary, when the input data is an open scene or a high-speed motion scene, the residual is close to zero, indicating that there is no target; when the input data is a target echo, the residual is the target energy after denoising. Target detection (CFAR) is performed on the residual result, and distance interpolation is performed. The appropriate antenna array element is selected for angle measurement to obtain the target position.

[0025] In steps S1 and S2, when extracting data from multiple range-dimensional sampling units in the radar close-range area of ​​each receiving channel, if the data volume is small, a lightweight network can be selected for training to facilitate on-chip deployment.

[0026] Step S5: Target real-time data denoising, detection, and recognition,

[0027] When the radar is operating normally, it collects 2DFFT results of the distance across each receiving channel in real-time and normal operation. This data includes data from the radar's close-range area (0-5 meters from the radar) and data from the long-range area (more than 5 meters from the radar). The long-range data is used for normal driving radar functions, while the close-range data is used for parking radar functions.

[0028] When the radar movement speed is higher than 2m / s, the radar close-range target detection module is not enabled for close-range target detection.

[0029] When the radar speed is lower than 2 m / s, the radar close-range target detection module is activated to perform close-range target detection. Close-range target detection includes: extracting data from multiple range-dimensional sampling units in the radar close-range area of ​​each receiving channel. At the same time, the real-time temperature of the radar RF chip is recorded and extracted, so that the extracted data forms the real-time target data. The real-time target data is input into the target detection network finally obtained in step S4. After passing through the denoising network, the target data of each channel is obtained. The data of each channel are summed to obtain a range Doppler map (RDMAP). Target detection is performed on the range Doppler map, and then the current frame target information is obtained through range interpolation, speed measurement and angle measurement (i.e., conventional signal processing is performed on the RDMAP: CFAR (target detection), speed measurement, angle measurement (DoA), and range interpolation to obtain target position information). Multiple frames of range Doppler maps (RDMAP) are accumulated, and three consecutive frames of range Doppler maps are extracted and input into the classification network to obtain the target category. At this time, the target category and the latest saved target information are output to facilitate system decision-making.

[0030] If no target is detected in three consecutive frames of range-Doppler maps within the radar's close-range target detection range, indicating that there is no target within the radar's observation range during parking, the target parameter information and range-Doppler map of the corresponding frames will be cleared.

[0031] It should be noted that the range-Doppler map is a two-dimensional spectrum map generated by 2DFFT, which reflects the joint distribution of range and velocity of the target.

[0032] Compared with the prior art, the present invention has the following technical effects:

[0033] This invention addresses the blind spot issue of millimeter-wave radar in close-range detection and proposes an automatic parking obstacle perception method based on millimeter-wave radar. Through data acquisition, network model construction, and model training, a denoising network and a classification network are ultimately generated. The denoising network outputs denoised target data based on the real-time input RDMAP and corresponding operating temperature. Through CFAR, velocity measurement, angle measurement, and distance interpolation, the target's position relative to the radar is calculated. By accumulating multiple frames of RDMAP, the classification network outputs the target category. The denoising network removes noise energy and improves target distance accuracy. Furthermore, during data set creation and real-time processing, the radar RF chip temperature is detected and used as one of the network model inputs, improving the robustness of the denoising network at different temperatures and preventing temperature-induced noise residue. When an obstacle is within the radar's observation range, the target category, position, and velocity information are output, facilitating system decision-making. When no obstacle is present, false alarms are avoided, preventing the system from making erroneous decisions.

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0036] Figure 1 It is a schematic flow chart of the method of the present invention;

[0037] Figure 2 2 is a schematic diagram showing how the DC noise energy of the second distance sampling unit changes with chip temperature in an embodiment of the present invention;

[0038] Figure 3 is a schematic diagram showing how target energy varies with chip temperature in an embodiment of the present invention;

[0039] Figure 4 is a schematic diagram of a denoising network in an embodiment of the present invention;

[0040] Figure 5 is a schematic diagram of a classification network in an embodiment of the present invention;

[0041] Figure 6 is a schematic diagram of an original RDMAP when there is no target in the second range sampling unit in an embodiment of the present invention;

[0042] Figure 7 is a schematic diagram of the RDMAP after denoising when there is no target in the second range sampling unit in an embodiment of the present invention;

[0043] Figure 8 This is a schematic diagram of an original RDMAP when the second range sampling unit has a target in an embodiment of the present invention;

[0044] Figure 9 is a schematic diagram of the RDMAP after denoising when there is a target in the second range sampling unit in an embodiment of the present invention;

[0045] Figure 10 3. It is a schematic diagram of the denoised CFAR result when there is a target in the second range sampling unit in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. The following description sets forth numerous specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0047] For automatic parking scenarios, this embodiment proposes an automatic parking obstacle perception method based on millimeter-wave radar to meet the needs of the intelligent driving system to accurately perceive obstacles around the vehicle body during parking, thereby achieving efficient and safe automatic parking.

[0048] like Figure 1 The specific steps are as follows:

[0049] (1) Untargeted dataset creation,

[0050] The system collects 2DFFT results of the radar's range for each receiving channel in various target-free scenarios (including various open spaces with a stationary radar, various open spaces with a low-speed radar movement, and high-speed radar movement (where the low-speed radar is less than 2 m / s and the high-speed radar is greater than 10 m / s). The system then extracts data from multiple range-dimensional sampling units within the radar's close-range region (the radar close-range region is defined as the range 0-5 meters from the radar, and this range can be adjusted based on the actual application) for each receiving channel. While collecting data, the radar RF chip temperature is recorded and extracted to ensure that the target-free dataset covers as many scenarios as possible and includes all possible operating temperatures (upper and lower limits) of the radar RF chip. Ultimately, this extracted data forms the target-free data sample.

[0051] In order to better illustrate that the energy of DC and low-frequency noise changes with the temperature of the radar RF chip, this embodiment extracts the data of the second distance sampling unit (the second distance sampling unit is in the distance dimension) in the radar close range area, such as Figure 2 As shown in FIG, it is a schematic diagram showing that the DC noise energy of the second distance sampling unit changes with the chip temperature of the radar RF chip. It can be concluded that the DC noise energy of the second distance sampling unit decreases with the increase of the chip temperature of the radar RF chip.

[0052] (2) Target dataset creation,

[0053] Various targets are set up in the radar's close-range area and categorized into strong, stationary, weak, and moving targets based on their characteristics. Multiple range angles are set for each target. The 2DFFT results of the range for each radar receiving channel are collected for the same target in both stationary and low-speed motion states (where the radar's speed is less than 2 m / s). Data from multiple range-dimensional sampling units in the radar's close-range area are extracted from each receiving channel. While collecting this data, the temperature of the radar's RF chip is recorded and retrieved. This extracted data forms target data samples, which are then labeled with the target category.

[0054] In order to better illustrate that the target energy changes with the temperature of the radar RF chip, this embodiment takes the target energy at 50m, such as Figure 3 As shown in FIG, it is a schematic diagram showing the change of target energy with the change of chip temperature. It can be concluded that the target energy at 50 m decreases as the chip temperature of the radar RF chip increases.

[0055] In summary, noise energy and target energy are both related to the operating temperature of the radar RF chip. Therefore, when collecting data, it is necessary to record the corresponding radar RF chip operating temperature and use the radar RF chip operating temperature as input at the same time to avoid missed detection and false detection caused by temperature changes.

[0056] (3) Network model construction,

[0057] First, a denoising network is constructed. The targetless data samples are used as training samples and input into the denoising network for training. The residual is output and the training iteration is continued until the residual is equal to zero, thereby obtaining a denoising network with a residual equal to zero (i.e., a lightweight denoising network).

[0058] In this embodiment, the 2DFFT results of the second range sampling unit (the second range sampling unit is in the range dimension) of the radar close-range area of ​​each receiving channel are taken as input, and arranged in the Doppler dimension and channel dimension to obtain a two-dimensional image. U-net is selected as the denoising network and pruned to finally obtain a lightweight denoising network, as shown in FIG. Figure 4 shown.

[0059] Then, a target detection network is constructed: the target data samples are used as training samples and input into the lightweight denoising network for training. The residual is output and the target detection is performed on the residual result to determine whether the target exists. The target detection position and target category are obtained. The training iteration is continued until the target detection position and target category correspond one-to-one with the set target position and target category, thereby obtaining the target detection network and classification network.

[0060] After the denoising network training is completed, the target detection network is constructed. After the denoising network, the real target echo can be obtained. Multiple frames are accumulated as the input of the target detection network. The target detection network is trained and the output result is the target category, which corresponds to the label obtained by classification based on the characteristics. In this embodiment, the range Doppler map (RDMAP) accumulated from 3 frames is used as input, the convolutional neural network is selected as the feature extraction network, and the fully connected layer is used as the classifier to obtain a fully connected neural network (i.e., classification network), as shown in the following figure. Figure 5 shown.

[0061] (4) Target real-time data denoising, detection, and recognition,

[0062] When the radar is operating normally, it collects 2DFFT results of the distance across each receiving channel in real-time and normal operation. This data includes data from the radar's close-range area (0-5 meters from the radar) and data from the long-range area (more than 5 meters from the radar). The long-range data is used for normal driving, while the close-range data is used for parking.

[0063] When the radar movement speed is higher than 2m / s, the radar close-range target detection module is not enabled for close-range target detection.

[0064] When the radar speed is lower than 2 m / s, the radar close-range target detection module is activated for close-range target detection. Close-range target detection includes: extracting data from multiple range-dimensional sampling units in the radar close-range area of ​​each receiving channel. At the same time, the real-time temperature of the radar RF chip is recorded and extracted, thereby forming the real-time target data. The real-time target data is input into the target detection network. After passing through the denoising network, the target data of each channel is obtained. The data of each channel is summed to obtain the range Doppler map (RDMAP). Target detection is performed on the range Doppler map, and then the current frame target information is obtained through range interpolation, speed measurement and angle measurement (that is, the RDMAP is subjected to conventional signal processing: CFAR (target detection), speed measurement, angle measurement (DoA), and range interpolation to obtain target position information). Multiple frames of range Doppler maps (RDMAP) are accumulated, and three consecutive frames of range Doppler maps are extracted and input into the classification network to obtain the target category. At this time, the target category and the latest saved target information are output to facilitate system decision-making.

[0065] If no target is detected in three consecutive frames of range-Doppler maps within the radar's close-range target detection range, indicating that there is no target within the radar's observation range during parking, the target parameter information and range-Doppler map of the corresponding frames will be cleared.

[0066] In this embodiment, the 2DFFT results of the second range sampling unit (the second range sampling unit is in the range dimension) of the radar close range area of ​​each receiving channel are taken as input, such as Figure 6 As shown, this is the original RDMAP schematic diagram when there is no target in the second range sampling unit; Figure 7 As shown in the figure, it is a schematic diagram of RDMAP after denoising when there is no target in the second distance sampling unit; Figure 8 As shown in FIG, the original RDMAP schematic diagram when the second distance sampling unit has a target, where the target is a pedestrian; Figure 9 As shown in FIG, this is a schematic diagram of the RDMAP after denoising when there is a target in the second distance sampling unit, where the target is a pedestrian; Figure 10 The figure shows the denoised CFAR result when there is a target in the second range sampling unit, where the target is a pedestrian.

[0067] In summary, the present invention proposes a millimeter-wave radar-based automatic parking obstacle perception method that uses a denoising network to remove DC signal energy. This method not only improves weak target detection performance but also enhances the ranging accuracy of close-range targets. Furthermore, the radar RF chip operating temperature is monitored during data set creation and real-time processing and used as one of the network inputs, enhancing the robustness of the denoising network at different temperatures and preventing temperature-induced noise residue. After denoising, target detection, velocity measurement, angle measurement, and range interpolation are performed to obtain target parameter information. Multi-frame energy distribution maps are accumulated and then classified using a classification network. Unlike conventional radar target classification, the present invention directly uses the Doppler frequency-time energy distribution map (i.e., the range-Doppler map) as the network input for target classification. This input contains more target feature information, significantly improving millimeter-wave radar's close-range target detection performance and ranging accuracy. Compared to ultrasonic radar, this method not only outputs highly accurate ranging results but also provides target position, altitude, and velocity information. In the present invention, only close-range targets are classified, and only the Doppler results of some distance-dimensional sampling units need to be selected as network input. Multi-frame accumulation is also used, which greatly reduces the amount of calculation and memory consumption.

[0068] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any person skilled in the art can utilize the methods and technical contents disclosed above to make many possible variations and modifications to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, or modify them into equivalent embodiments with equivalent variations. Therefore, any equivalent variations made in accordance with the shape, structure, and principles of the present invention without departing from the content of the technical solutions of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for automatic parking obstacle perception based on millimeter wave radar, characterized in that: The following steps are involved: Step S1: Untargeted dataset creation, Collect the 2DFFT results of the radar's receiving channel distance in various target-free scenarios, extract the data of multiple range-dimensional sampling units in the radar's close-range area of ​​each receiving channel, and simultaneously record and extract the temperature of the radar RF chip. The extracted data thus forms target-free data samples. Step S2: Target dataset creation, Various targets are set in the radar's close-range area, with multiple distance angles set for each target. The 2DFFT results of the distances of each radar receiving channel are collected for the same target in the radar's motion state. The data of multiple range-dimensional sampling units in the radar's close-range area of ​​each receiving channel are extracted. At the same time, the temperature of the radar RF chip is recorded and extracted. The extracted data forms the target data sample, and the set target is labeled with the label corresponding to the target category. Step S3: Construct a denoising network, use the target-free data samples from step S1 as training samples, and continue training iterations to obtain a denoising network with a residual equal to zero; Step S4: The target data samples in step S2 are used as training samples and input into the denoising network finally obtained in step S3. The training iteration is continued until the target detection position and target category correspond to the target position and target category set in step S2, thereby obtaining the target detection network and classification network. Step S5: Target real-time data denoising, detection, and recognition, The 2DFFT results of the distances of each receiving channel of the radar are collected when the radar is working normally in real time. The data of multiple distance dimension sampling units in the radar close-range area of ​​each receiving channel are extracted. At the same time, the real-time temperature of the radar RF chip is recorded and extracted, so that the extracted data forms the real-time target data. The real-time target data is input into the target detection network finally obtained in step S4 to obtain the current frame target information, which is input into the classification network to obtain the target category.

2. The automatic parking obstacle perception method based on millimeter wave radar according to claim 1, characterized in that: In step S1 , the target-free scene includes various open scenes with stationary radars, various open scenes with low-speed radar movement, and high-speed radar movement scenes.

3. The automatic parking obstacle perception method based on millimeter wave radar according to claim 1, characterized in that: In step S2, the radar motion state is divided into a stationary state and a low-speed motion state, and the speed of the radar in the low-speed motion state is lower than 2 m / s.

4. The automatic parking obstacle perception method based on millimeter wave radar according to claim 1, characterized in that: In step S2, after various targets are set in the radar close range area, the targets are divided into stationary strong targets, stationary weak targets and moving targets according to their characteristics.

5. The automatic parking obstacle perception method based on millimeter wave radar according to claim 1, characterized in that: In step S3, the process of obtaining a denoising network with a residual equal to zero includes: constructing a denoising network, inputting the target-free data sample of step S1 into the denoising network as a training sample, performing training, and outputting the residual, and continuously iterating the training until the residual is equal to zero, thereby obtaining a denoising network with a residual equal to zero.

6. The automatic parking obstacle perception method based on millimeter wave radar according to claim 1, characterized in that: In step S4, the process of obtaining the target detection position and its target category includes: using the target data sample of step S2 as a training sample, inputting it into the denoising network finally obtained in step S3, performing training, and outputting the residual, performing target detection on the residual result, judging whether the target exists, and obtaining the target detection position and target category.

7. The automatic parking obstacle perception method based on millimeter wave radar according to claim 1, characterized in that: In step S5, the process of obtaining the target information and target category of the current frame includes: inputting the real-time target data into the target detection network finally obtained in step S4 to obtain the range Doppler map, performing target detection on the range Doppler map, and then obtaining the target information of the current frame through distance interpolation, speed measurement and angle measurement, taking out multiple consecutive frames of range Doppler maps, and inputting them into the classification network to obtain the target category.

8. The automatic parking obstacle perception method based on millimeter wave radar according to claim 1, characterized in that: In step S5, when the radar movement speed is higher than 2 m / s, the radar close-range target detection module is not enabled to perform close-range target detection; When the radar movement speed is lower than 2m / s, the radar close-range target detection module is enabled to perform close-range target detection. Close-range target detection includes: extracting data from multiple range-dimensional sampling units in the radar close-range area of ​​each receiving channel, and at the same time, recording and extracting the real-time temperature of the radar RF chip, so that the extracted data forms the target real-time data, and inputting the target real-time data into the target detection network finally obtained in step S4 to obtain a range Doppler map, performing target detection on the range Doppler map, and then obtaining the current frame target information through distance interpolation, speed measurement and angle measurement, extracting multiple consecutive frames of range Doppler maps, and inputting them into the classification network to obtain the target category.

9. The automatic parking obstacle perception method based on millimeter wave radar according to claim 8, characterized in that: If no target is detected in three consecutive range-Doppler image frames within the radar's close-range target detection range, the target parameter information and range-Doppler image of the corresponding frames will be cleared.

10. The automatic parking obstacle perception method based on millimeter wave radar according to claim 4, characterized in that: In step S2, the moving target is set as a target whose relative movement speed exceeds one speed resolution unit of the radar; the stationary target is set as a target whose relative movement speed does not exceed one speed resolution unit of the radar; the stationary strong target is set as a strong scattering stationary target, which includes metal and walls; the stationary weak target is set as a weak scattering stationary target, which includes cones and green belts.

Images