A rain interference detection method and device based on a neural network

Abstract

The present invention discloses a rain interference detection method and device based on a neural network. Among them, the rain interference detection method based on a neural network includes the steps of: obtaining vehicle parameter information and point cloud data scanned by a radar; the vehicle parameter information includes the current speed of the vehicle; based on a preset detection threshold, selecting target point clouds that meet the detection threshold from the point cloud data; the detection threshold includes a spatial position area; inputting the selected target point clouds and the vehicle parameter information into a trained rain detection neural network model; and identifying and outputting whether each target point cloud is a rain interference point cloud. By adopting the scheme of judging point cloud features based on a neural network in this application, rain interferences such as close-range raindrops and splashing water can be detected, with fast recognition speed, high recognition accuracy, and low manual consumption.

Description

Technical Field

[0001] The present invention relates to the field of rain interference detection, and in particular to a rain interference detection method and device based on a neural network. Background Art

[0002] To meet the requirements of current automotive technology, a reverse radar, although it has the ability to detect non-metallic targets at close range, however, in actual use, there are some problems, that is, when it is raining, the reverse radar will receive radar waves reflected by falling raindrops or splashing accumulated water, as well as radar waves reflected by water droplets attached to the bumper. All of these may cause false alarms and inaccurate detections, causing driving troubles for the driver.

[0003] Rain interference detection is a high-frequency research scenario for millimeter-wave radars. Currently, a method based on the CFAR (Constant False Alarm Rate Detector) threshold is often used for determination. CFAR generally judges based on the energy spectrum of radar signals.

[0004] This CFAR-threshold-based determination method, on the one hand, is difficult to optimize in mass-produced vehicles through the threshold determination method (such as being affected by the bumper), the calculation process is complex, and the amount of calculation is relatively large; on the other hand, it is also easy to be confused with close-range targets. Summary of the Invention

[0005] In view of the above-mentioned defects of the prior art, the present invention provides a rain interference detection method and device based on a neural network, which solves the technical problem of false alarms and inaccurate detections caused by rain interference in reverse radars, and overcomes the defects of the prior art such as large calculation amount and easy confusion between rain interferences such as falling raindrops and splashing water flowers and close-range targets.

[0006] To achieve the above object, the present invention provides a rain interference detection method based on a neural network, and the method includes the following steps:

[0007] Obtain vehicle parameter information and point cloud data scanned by the radar; the vehicle parameter information includes the current speed of the vehicle;

[0008] Based on a preset detection threshold, select target point clouds that meet the detection threshold from the point cloud data; the detection threshold includes a spatial position area;

[0009] Input the selected target point clouds and the vehicle parameter information into a trained rain detection neural network model;

[0010] Identify and output whether each target point cloud is a rain interference point cloud.

[0011] Further, after identifying whether each target point cloud is a rain interference point cloud, the following steps are also included:

[0012] Count the number of point clouds identified as rain interference point clouds;

[0013] Determine whether the proportion of the number of the rain interference point clouds in the number of the target point clouds reaches a preset proportion threshold. If so, send a prompt of detecting rain interference to the radar.

[0014] Further, the detection threshold also includes: a Doppler velocity interval and / or a reflection energy interval; based on the preset detection threshold, selecting target point clouds that meet the detection threshold from the point cloud data specifically includes:

[0015] Select the point clouds falling into the preset spatial position area from the point cloud data as the first point clouds;

[0016] Screen the point clouds in the preset Doppler velocity interval and / or the preset reflection energy interval from the first point clouds as the target point clouds.

[0017] Further, the steps of training a rain detection neural network model are also included, specifically including:

[0018] Collect the point cloud data and vehicle parameter information of the scene with rain interference;

[0019] Based on the preset spatial position area, select the point clouds falling into the preset spatial position area from the point cloud data of the scene with rain interference as the initial training samples;

[0020] Identify whether the scene corresponding to the initial training sample is a pure rain interference scene;

[0021] When it is determined that the scene corresponding to the initial training sample is a pure rain interference scene, label each target point cloud in the initial training sample, and store the labeled initial training sample as a positive sample together with the corresponding vehicle's own speed in the training sample set;

[0022] Train the initial neural network model through the training sample set until a rain detection neural network model that meets the requirements is obtained.

[0023] Further, before training the initial neural network model through the training sample set, the following steps are also included:

[0024] Collect the point cloud data and vehicle parameter information of the scene without rain interference;

[0025] Based on the preset spatial position area, select the point clouds falling into the preset spatial position area from the point cloud data of the scene without rain interference as the negative samples for training;

[0026] Label each point cloud in the negative sample, and store the labeled negative sample and the corresponding vehicle parameter information in the training sample set.

[0027] Further, the rain detection neural network model is an MLP neural network, and a spatial variation network is further provided before the input layer of the MLP neural network model, which is used to extract features with spatial invariance from the input information as the input of the MLP neural network.

[0028] Further, the vehicle parameter information further includes: the current yaw rate of the vehicle, and / or the vehicle model.

[0029] Before selecting the target point cloud that meets the detection threshold from the point cloud data based on a preset detection threshold, it further includes:

[0030] Retrieve the corresponding preset detection threshold according to the parameter information of the vehicle.

[0031] On the other hand, the present application also discloses a rain interference detection device based on a neural network, including:

[0032] An acquisition module, configured to acquire vehicle parameter information and point cloud data scanned by a radar;

[0033] A screening module, configured to select a target point cloud that meets the detection threshold from the point cloud data based on a preset detection threshold; the detection threshold includes a spatial position area;

[0034] An input module, configured to input the selected target point cloud and the vehicle parameter information into a trained rain detection neural network model;

[0035] An identification module, configured to identify and output whether each target point cloud is a rain interference point cloud through the rain detection neural network model.

[0036] Further, the detection threshold further includes: a Doppler velocity interval, and / or a reflection energy interval; the screening module specifically includes: a position screening sub-module, a velocity screening sub-module, and an energy screening sub-module; where:

[0037] The position screening sub-module is configured to select the point cloud that falls into the spatial position area from the point cloud data as the first point cloud;

[0038] The velocity screening sub-module is configured to screen the second point cloud that is within a preset Doppler velocity interval from the first point cloud;

[0039] The energy screening sub-module is configured to screen the point cloud that is within a preset reflection energy interval from the second point cloud as the target point cloud.

[0040] Furthermore, the rain detection neural network model is an MLP neural network, and a spatial transformation network is further provided before the input layer of the MLP neural network model, which is used to extract spatially invariant features from the input information as the input of the MLP neural network.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) This application is based on point clouds for feature extraction and judgment, rather than spectral features. If spectral features are used, the processes of obtaining and extracting spectral features are cumbersome, and the computational complexity is relatively higher, far less convenient than this application. The solution of this application can detect rainwater interference such as raindrops at close range and splashing water, and can improve the calculation speed.

[0043] (2) This application judges point cloud features based on an MLP neural network, with a fast training speed and low labor consumption. During training, positive and negative samples are used for training, and in the rainwater interference scenario, non-simple rainwater interference scenarios are screened out, so that the target point clouds can be batch-labeled, greatly improving the labeling speed of the target point clouds.

[0044] (3) The rainwater interference detection method based on a neural network in this application can be applied to different vehicle models. Different vehicle models have different radar installation heights, and the corresponding preset detection thresholds will also be different. For different vehicle models, only different detection thresholds need to be preset to screen the target point clouds. Of course, the trained rainwater interference neural network model can also be trained into a rain detection neural network model for a certain vehicle model during training, or can be trained into a neural network model that can identify rainwater interference detection for different vehicle models through training samples of different vehicle models during training.

[0045] (4) Optimize the MLP input based on STN to obtain spatial invariance, reduce the number of layers of the MLP, improve the recognition speed and accuracy, reduce the computational amount, and have better adaptability to different vehicle models.

[0046] The following will further illustrate the concept, specific structure and technical effects of the present invention with reference to the drawings, so as to fully understand the purpose, features and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flowchart of the rainwater interference detection method based on a neural network according to Embodiment 1 of the present invention;

[0048] Figure 2 is a flowchart of the rainwater interference detection method based on a neural network according to Embodiment 2 of the present invention;

[0049] Figure 3It is the training flowchart of the rain detection neural network model in the fourth embodiment of the present invention;

[0050] Figure 4 It is the logical topology diagram of the MLP neural network in the fifth embodiment of the present invention;

[0051] Figure 5 It is the connection schematic diagram of the STN network and the MLP neural network in the fifth embodiment of the present invention;

[0052] Figure 6 It is the structural block diagram of the rain interference detection device based on the neural network in the sixth embodiment of the present invention;

[0053] Figure 7 It is the working process schematic diagram of the rain interference detection device in the sixth embodiment of the present invention. Detailed implementation manners

[0054] The following introduces multiple preferred embodiments of the present invention with reference to the accompanying drawings of the specification to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0055] Embodiment 1

[0056] This embodiment provides a rain interference detection method based on a neural network, as Figure 1 shown, the method includes the following steps:

[0057] S101, obtaining vehicle parameter information and point cloud data scanned by a radar; the vehicle parameter information includes the current speed of the vehicle;

[0058] Specifically, the current vehicle speed can be obtained through a radar sensor, or can be obtained through other sensors of the vehicle itself and then transmitted over. In addition, the radar installed at the rear of the vehicle will continuously collect and scan the point cloud data of the current scene. Generally, for example, after the radar receives the original data of the scanned scene, it will process the received original data, and the processing methods include FFT, CFAR threshold detection, array antenna angle detection algorithm, etc. Finally, the point cloud data information of the radar is obtained and stored in the system.

[0059] S102, selecting target point clouds that meet the detection threshold from the point cloud data based on a preset detection threshold; the detection threshold includes a spatial position area;

[0060] Specifically, the preset detection threshold generally includes the spatial area relative to the vehicle where falling raindrops and splashing water (collectively referred to as rain interference) may appear. Generally, the establishment of this detection threshold needs to be obtained through multiple experimental measurements. Moreover, for different vehicle models, or rather, different radar installation heights, the corresponding values of the detection threshold will also be different. If different vehicle models are to be applicable, different detection thresholds corresponding to different vehicle models can be preset in advance. The vehicle parameter information obtained in the early stage can include not only the current vehicle speed information but also the basic information of the vehicle itself, such as the vehicle model, so as to facilitate the selection of different detection thresholds according to the vehicle parameter information in the later stage to screen the target point cloud. The establishment and screening of the detection threshold can extract and screen the point cloud in the spatial area where rain interference may appear, so that only the screened target point cloud needs to be identified and processed, greatly reducing the workload of subsequent processing and also improving the recognition speed.

[0061] S103, input the selected target point cloud and the vehicle parameter information into the trained rain detection neural network model;

[0062] Specifically, input the target point cloud obtained after the preliminary screening process and the vehicle parameter information into the trained rain detection neural network model for recognition and processing. The rain detection neural network model here has been trained in advance and can be used to detect whether the target point cloud is a rain interference point cloud.

[0063] In addition, there are two implementation methods for using the selected target point cloud and vehicle parameter information as the input of the rain detection neural network model. One is to directly input the target point cloud and vehicle motion data without processing into the rain detection neural network model. Another implementation method is to first perform fusion processing on the vehicle motion data and integrate it into the screened target point cloud, so that the target point cloud carries the vehicle motion data, and finally input the target point cloud map integrated with the vehicle motion data into the rain detection neural network model for recognition and detection.

[0064] S104, identify and output whether each target point cloud is a rain interference point cloud.

[0065] Specifically, through the rain detection neural network model of this embodiment, it can be identified and detected whether each target point cloud is an interference point cloud formed by rain interference.

[0066] This embodiment is based on point clouds for feature extraction and judgment, rather than spectral features. If spectral features are used, the processes of obtaining spectral features and extracting spectral features are cumbersome, and the calculation amount is relatively larger, far less convenient than this application. The solution of this application can detect rain interference such as close-range raindrops and splashing water, and can improve the calculation speed.

[0067] Preferably, after obtaining the recognition results of each target point cloud, the results can be further statistically analyzed to further determine whether the current scene is a rain interference scene, thus avoiding false alarms. Optionally, based on the above embodiment, the rain interference detection method based on a neural network may further include the following steps:

[0068] S105, count the number of point clouds recognized as rain interference point clouds;

[0069] S106, determine whether the proportion of the number of the rain interference point clouds in the number of the target point clouds reaches a preset proportion threshold. If so, enter step S107;

[0070] S107, send a prompt of detecting rain interference to the radar.

[0071] Specifically, for example, if the counted proportion of the number of point clouds recognized as rain interference clouds reaches more than 10%, it can be determined that the currently collected point cloud scene is a rain interference scene, and then a prompt of detecting rain interference will be sent to the radar, thus avoiding false alarms of the radar triggered by rain interference.

[0072] Embodiment 2

[0073] Based on the above embodiment, in this embodiment, in addition to the preset spatial position area where rain interference is located, a Doppler velocity interval and a reflection energy interval are also set for the preset detection threshold. Specifically, the rain interference detection method based on a neural network in this embodiment is as Figure 2 shown and includes the following steps:

[0074] S201, obtain vehicle parameter information and point cloud data scanned by the radar; the vehicle parameter information includes the current speed of the vehicle;

[0075] S202, select the point clouds falling into the preset spatial position area from the point cloud data as the first point clouds;

[0076] Specifically, in the preset detection threshold, the spatial position area is set as the 3D Cartesian coordinate system [x, y, z] based on the center of the vehicle itself, which is the spatial position area where rain interference (water splash and falling raindrops) is most likely to be located.

[0077] S203, screen out the point clouds in the preset Doppler velocity interval and / or the preset reflection energy interval from the first point clouds as the target point clouds;

[0078] When raindrops fall, they are affected by gravity and air resistance. When passing in front of the radar from top to bottom, they will produce energy reflections within a certain range of positive and negative values centered on 0 bin, forming a point cloud. Through multiple experimental measurements, the range of reflected energy caused by the falling rain can be obtained. The preset Doppler velocity range is set as the range where the radial relative velocity between the water splash and the falling raindrops and the radar is located, that is, the Doppler velocity range. The preset reflected energy range is also configured as the range where the reflected energy values of the water splash and the falling raindrops are located in the preset spatial position area, that is, the reflected energy range. Generally, considering the current radar hardware configuration, the Doppler velocity and reflected energy of each point cloud can be directly obtained by the radar.

[0079] S204, input the selected target point cloud and the vehicle parameter information into the trained rain detection neural network model;

[0080] S205, identify and output whether each target point cloud is a rain interference point cloud.

[0081] In this embodiment, through the preset spatial position area, the spatial position area where rain interference is most likely to occur can be extracted, and the point clouds in other places do not need to be processed, reducing the amount of data processing and recognition. In addition, in this embodiment, the point clouds with Doppler velocity in the preset Doppler velocity and / or reflected energy in the preset reflected energy range are further screened from the point clouds falling into the preset spatial position area as target point clouds, and only the finally screened target point clouds are used as input parameters. In this way, the irrelevant point clouds are further reduced, the amount of feature extraction for subsequent recognition is reduced, and the recognition efficiency and accuracy are improved.

[0082] Embodiment III

[0083] In this embodiment, in addition to the current vehicle speed, the vehicle parameter information also includes the current vehicle yaw rate and / or vehicle type. We take the vehicle parameter information including the current vehicle speed, the current vehicle yaw rate and the vehicle type as an example. The neural network-based rain interference detection method in this embodiment specifically includes:

[0084] S301, obtain the vehicle parameter information and the point cloud data scanned by the radar; the vehicle parameter information includes the current vehicle speed, the current vehicle yaw rate, and the vehicle type;

[0085] S302, retrieve the corresponding spatial position area, Doppler velocity range, and reflected energy range according to the vehicle parameter information;

[0086] For the same vehicle model, the settings of the spatial position area are generally the same. However, if the vehicle models are different, the installation heights of the radars at the rear of the vehicle may also be different, so the settings of the spatial position area will also be different. In addition, the Doppler velocity range and the reflection energy range will also change with the vehicle speed, yaw rate, etc. Therefore, if the set detection threshold is relatively fine, it will vary depending on the vehicle's parameter information. Taking water splash as an example, specifically, the accumulated water is affected by the traction force of the tire on it and the air resistance, and parabolic splashes are generated at the rear of the vehicle, forming splashing water. The different shapes and operating states of the rear parts of different vehicle models cause complex changes in the splashing water. Considering from the perspective of the rear radar, when the splashing water drifts to the front of the radar, energy reflections within a certain range of positive and negative values centered on 0 bin will be generated. The reflection energy range will increase as the speed of the vehicle itself increases, that is, as the speed of the tire pulling the accumulated water increases.

[0087] S303, based on the retrieved spatial position area, select the first point cloud falling into the spatial position area from the point cloud data;

[0088] S304, based on the retrieved Doppler velocity range and reflection energy range, screen the target point cloud that meets the Doppler velocity range and the reflection energy range from the first point cloud.

[0089] S305, input the selected target point cloud and the vehicle parameter information into the trained rain detection neural network model;

[0090] Specifically, taking the screened target point cloud and vehicle parameter information together as the input parameters of the rain detection neural network model can efficiently and quickly extract classification features that are difficult for humans to obtain, so as to flexibly cope with the different characteristics of rain interference false targets caused by different vehicle models.

[0091] S306, identify and output whether each target point cloud is a rain interference point cloud.

[0092] The rain interference detection method based on neural network in this application can be applied to different vehicle models. If the vehicle models are different and the installation heights of the vehicle radars are different, the corresponding preset detection thresholds will also be different. For different vehicle models, only need to preset different detection thresholds to screen the target point cloud, reducing the workload of subsequent feature extraction. And when identifying, it fuses the target point cloud information and vehicle parameter information, thus improving the accuracy of identification.

[0093] Embodiment 4

[0094] Based on any of the above embodiments, this embodiment focuses on elaborating the training process of the rain detection neural network model. During training, the point cloud data of the pure rain interference scenario can be used as positive samples for training. More preferably, the point cloud data positive samples of the pure rain interference scenario can be combined with the point cloud data negative samples of the non-rain interference scenario for joint training. Specifically, as Figure 3 shown, the training method of the rain detection neural network model includes the following steps:

[0095] S001, collect the point cloud data of the scene with rain interference and vehicle parameter information;

[0096] Specifically, after the radar receives the original scanned data, the original scanned data is processed to obtain point cloud data. In this embodiment, the method for processing the original data includes at least one of FFT (fast Fourier transform), CFAR threshold detection method, or array antenna angle detection method.

[0097] S002, based on a preset spatial position area, select the point cloud that falls into the preset spatial position area from the point cloud data of the scene with rain interference as the initial training sample;

[0098] Specifically, the preset spatial position area is generally set as the area where rain interference (falling raindrops and splashing water) may occur. For different vehicle models, the height of the radar installed at the rear of the vehicle may also be different. Therefore, different vehicle models can preset different spatial position areas. This spatial position area is generally obtained through multiple tests and calculations.

[0099] S003, identify whether the scene corresponding to the initial training sample is a pure rain interference scene; if so, proceed to step S004;

[0100] Specifically, among the collected rain interference scenes, some may be pure rain interference scenes, and some may be scenes with both rain interference and other obstacle interferences. In this embodiment, for the convenience of subsequent unified annotation, reducing the annotation workload, and being able to output the recognition results of each target point cloud in the future, therefore, in this embodiment, the point cloud data of the pure rain interference scene will be selected, and the point cloud data of the non-pure rain interference scene will be removed. The identification method can be to manually identify the collected scene video, or other identification methods can also be used. For example, through the Doppler velocities and reflection energies of the collected scene point clouds, etc., comprehensively analyze and judge whether the scene point cloud is a pure rain interference scene point cloud. Taking the reflection energy as an example, the reflection energy of a nearby obstacle is generally relatively large. Therefore, if there is a reflection energy in the collected rain interference scene that exceeds the preset reflection energy range, it can be determined that the rain interference scene is a non-pure rain interference scene.

[0101] S004, label each target point cloud in the initial training sample, and store the labeled initial training sample as a positive sample along with the corresponding vehicle parameter information in the training sample set;

[0102] Specifically, since the non - simple rain interference scenarios have been eliminated in step S003, the workload in this labeling step is greatly reduced. After selection, the remaining point cloud data are all from simple rain interference scenarios. Therefore, when labeling, each target point cloud can be uniformly labeled as a true point cloud (rain interference point cloud).

[0103] S005, collect point cloud data and vehicle parameter information for scenarios without rain interference;

[0104] S006, based on a preset spatial position area, select the point cloud that falls into the preset spatial position area from the point cloud data of the scenario without rain interference as a negative sample for training;

[0105] S007, label each point cloud in the negative sample, and store the labeled negative sample along with the corresponding vehicle parameter information in the training sample set;

[0106] S008, train the initial neural network model with the training sample set until a rain detection neural network model that meets the requirements is obtained.

[0107] For different vehicle models and different application requirements, as long as the positive and negative samples for training data are reasonably prepared, the optimal parameters for distinguishing rain interference can be automatically obtained through the backpropagation method, obtaining a more accurate recognition ability than traditional methods without the need to increase additional design work. After the model training is completed, within the spatial range where there may be rain interference, taking the relevant parameters of the point cloud and the information of the vehicle itself as input information can effectively distinguish the different characteristics of rain interference from close - range obstacles under different vehicle models.

[0108] Embodiment Five

[0109] In any of the above - mentioned embodiments, the rain detection neural network model is an MLP neural network. Figure 4 is the logical topology diagram of the MLP neural network, including an input layer 1, at least one hidden layer 2, and an output layer 3; the hidden layer 2 calculates the information received by the input layer 1; the output layer 3 outputs the judgment result calculated by the hidden layer 2. There is a full connection between the input layer 1, the hidden layer 2, and the output layer 3; each node in the hidden layer 2 has a weight, an activation function, and a bias.

[0110] The MLP neural network determines whether the first point cloud is the rain interference through the forward propagation calculation of the hidden layer 2; the weights, activation functions, and biases of each node in the hidden layer 2 are obtained by training based on the actual data through the backpropagation method.

[0111] Input the information of the input layer 1, including the vehicle's own speed, vehicle yaw rate (angular velocity), spatial position of the target point cloud, reflected energy of the target point cloud, or Doppler velocity of the target point cloud.

[0112] Among them, the spatial position of the target point cloud is set as the 3D Cartesian coordinate system [x, y, z] based on the vehicle center; the reflected energy of the target point cloud is set as the radial relative velocity between the reflecting object and the radar. In addition, considering the current radar hardware configuration, the Doppler velocity of the target point cloud can be directly provided by the radar processing module.

[0113] Take the spatial range that may be affected by rain obtained through a large number of tests as the spatial position area in the detection threshold. Obtain the target point cloud through the screening of the spatial position area, and then use the relevant parameters of the target point cloud and vehicle parameters (such as vehicle speed) as input information, which can effectively distinguish the different characteristics of rain interference relative to close-range obstacles under different vehicle models.

[0114] In this embodiment, based on the above MLP neural network, an STN network (Spatial Transformer Networks) is set before the input of the MLP. Specifically, as shown in the figure, a spatial transformation network (STN network) is also set before the input layer of the MLP neural network model, which is used to extract features with spatial invariance from the input information as the input of the MLP neural network.

[0115] As Figure 5 shown, the STN network includes a local network Local net and a transformation matrix T; the local network Localnet is a fully connected MLP model, and the transformation matrix T is the transformation matrix of the input X, which transforms the input X into X' with spatial invariance. Compared with only setting the MLP neural network model, the information originally input to the input layer needs to be processed by the local network Local net and the transformation matrix T in sequence before being input to the input layer.

[0116] Specifically, at different vehicle speeds and different spatial positions, the point cloud information of the same rain interference is different, such as the data of the spatial position of the point cloud, the vehicle angle, or the Doppler velocity of the point cloud. In order to record and eliminate the influence of the same judgment result but different input information, the traditional MLP neural network requires a relatively large depth (i.e., the number of network layers). The STN network is placed in front of the MLP neural network, transforming the original input features into standardized features with spatial invariance, and solving the problem of spatial invariance of the input information. Under the condition of the same overall effect, the attention mechanism introduced by the STN network can maintain the features of the feature map in space, reduce the number of network layers of the MLP, and reduce the computational amount and the storage capacity of the network model.

[0117] Embodiment VI

[0118] Based on the same technical concept, this embodiment provides a rain interference detection device based on a neural network, as Figure 6 shown, including:

[0119] An acquisition module 100, configured to acquire vehicle parameter information and point cloud data scanned by a radar; the vehicle parameter information includes the current speed of the vehicle;

[0120] A screening module 200, configured to select target point clouds that meet the detection threshold from the point cloud data based on a preset detection threshold; the detection threshold includes a spatial position area;

[0121] An input module 300, configured to input the selected target point clouds and the vehicle parameter information into a trained rain detection neural network model;

[0122] An identification module 400, configured to identify and output whether each target point cloud is a rain interference point cloud through the rain detection neural network model.

[0123] Preferably, the rain interference detection module of this embodiment may further include:

[0124] A statistics module, configured to count the number of point clouds identified as rain interference point clouds;

[0125] A prompt module, configured to determine whether the proportion of the number of rain interference point clouds in the number of target point clouds reaches a preset proportion threshold, and if so, send a prompt of detecting rain interference to the radar.

[0126] Figure 7 Shown is a schematic diagram of the detection and identification working process of the rain interference detection device disclosed in this embodiment.

[0127] S601. The acquisition module collects point cloud data through a radar system and acquires vehicle parameter information;

[0128] S602. The screening module detects based on a preset detection threshold for rain interference whether there is any point cloud falling within the detection threshold; if so, go to step S603; otherwise, go to step S605;

[0129] S603. The input module inputs the target point clouds falling within the detection threshold into a trained MLP neural network (rain detection neural network model);

[0130] S604. The identification module determines whether each point cloud falling within the detection threshold is rain interference; if so, go to step S606; otherwise, go to step S605;

[0131] S605. The recognition module labels the point cloud corresponding to a normal target, i.e., a false point cloud.

[0132] S606. The recognition module labels the point cloud for which the S604 judgment result is true as corresponding to rain interference (true point cloud), and performs specific processing on it; the specific processing includes labeling the category or deletion operation.

[0133] S607. The statistics module identifies the number of point clouds for rain interference (true point clouds).

[0134] S608. When the proportion of the number of the rain interference point clouds in the number of the target point clouds reaches a preset proportion threshold, the prompt module gives a prompt of detecting rain interference to the radar system.

[0135] Embodiment Seven

[0136] Based on the above Embodiment Six, the detection threshold further includes: a Doppler velocity range, and / or a reflection energy range; the screening module specifically includes: a position screening sub-module, and a velocity screening sub-module and / or an energy screening sub-module.

[0137] Taking the screening module including a position screening sub-module, a velocity screening sub-module, and an energy screening sub-module as an example, where:

[0138] The position screening sub-module is used to select the point cloud falling into the spatial position area from the point cloud data as the first point cloud.

[0139] The velocity screening sub-module is used to screen the second point cloud in a preset Doppler velocity range from the first point cloud.

[0140] The energy screening sub-module is used to screen the point cloud in a preset reflection energy range from the first point cloud or the second point cloud as the target point cloud.

[0141] In this way, through the screening of the Doppler velocity range and the reflection energy range, the point clouds without rain interference can be further removed, reducing the workload of subsequent input recognition. Of course, the target point cloud can also be obtained only through position screening and velocity screening, or only through position screening and energy screening.

[0142] In any of the above device embodiments, the rain detection neural network model is an MLP neural network. In the existing method, the energy spectrum is used as the recognition object, and the detection threshold is generally obtained through statistical methods, and multiple detection thresholds are used as the judgment means. Compared with the existing method, the method disclosed in this embodiment is based on the MLP neural network and the information of the point cloud, and can process falling raindrops and splashing water flowers simultaneously in one calculation, and can distinguish rain interference from pedestrians, two-wheel vehicles, and vehicles with small reflected energy or fast speed, reducing the secondary development workload of the MLP neural network when switching the installed vehicle model or radars with similar characteristics.

[0143] Preferably, a spatial transformation network is further provided before the input layer of the MLP neural network model, which is used to extract features with spatial invariance from the input information as the input of the MLP neural network.

[0144] Specifically, at different self-vehicle speeds and different spatial positions, the point cloud information of the same rain interference is different, such as the data of the point cloud spatial position, the self-vehicle angle, or the point cloud Doppler velocity. In order to record and eliminate the influence of different input information with the same judgment result, the traditional MLP neural network needs to consume more depth (i.e., the number of network layers). The STN network is placed in front of the MLP neural network, which transforms the original input features into standardized features with spatial invariance, and solves the problem of spatial invariance of the input information. In the case of the same overall effect, the attention mechanism introduced by the STN network can maintain the features of the feature map in space, reduce the number of network layers of the MLP, and reduce the calculation amount and the storage capacity of the network model.

[0145] In addition, with the popularization of the application of high-performance neural network dedicated processing chips / processors, the MLP neural network algorithm is more suitable for the development trend of the industry.

[0146] The device embodiment of the present application corresponds to the method embodiment of the present invention. The technical details of the method embodiment of the present invention are equally applicable to the device embodiment of the present invention. To avoid repetition, they will not be elaborated here.

[0147] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative labor. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the existing technology shall fall within the protection scope determined by the claims.

Claims

1. A rain interference detection method based on a neural network, characterized in that, the method comprises the following steps: Obtain vehicle parameter information and point cloud data scanned by a radar; the vehicle parameter information includes the current speed of the vehicle; Based on a preset detection threshold, select target point clouds that meet the detection threshold from the point cloud data; the detection threshold includes a spatial position area; Input the selected target point clouds and the vehicle parameter information into a trained rain detection neural network model; Identify and output whether each target point cloud is a rain interference point cloud; The detection threshold further includes: a Doppler velocity interval, and / or a reflection energy interval; based on the preset detection threshold, selecting target point clouds that meet the detection threshold from the point cloud data specifically includes: Select point clouds that fall into the preset spatial position area from the point cloud data as the first point clouds; Screen point clouds that are in the preset Doppler velocity interval and / or the preset reflection energy interval from the first point clouds as target point clouds; The method further includes the step of training a rain detection neural network model, specifically including: Collect point cloud data and vehicle parameter information of a scene with rain interference; Based on a preset spatial position area, select point clouds that fall into the preset spatial position area from the point cloud data of the scene with rain interference as initial training samples; Identify whether the scene corresponding to the initial training sample is a pure rain interference scene; When it is determined that the scene corresponding to the initial training sample is a pure rain interference scene, label each target point cloud in the initial training sample, and store the labeled initial training sample as a positive sample together with the corresponding vehicle parameter information in a training sample set; Train an initial neural network model through the training sample set until a rain detection neural network model that meets the requirements is obtained.

2. The rain interference detection method based on a neural network according to claim 1, characterized in that, after identifying whether each target point cloud is a rain interference point cloud, it further includes: Count the number of point clouds identified as rain interference point clouds; Judge whether the proportion of the rain interference point clouds in the number of target point clouds reaches a preset proportion threshold, and if so, send a prompt of detecting rain interference to the radar.

3. The rain interference detection method based on a neural network according to claim 2, characterized in that, before training the initial neural network model through the training sample set, it further includes: Collect point cloud data and vehicle parameter information of a scene without rain interference; Based on a preset spatial position area, select point clouds that fall into the preset spatial position area from the point cloud data of the scene without rain interference as negative samples for training; Label each point cloud in the negative sample, and store the labeled negative sample together with the corresponding vehicle parameter information in a training sample set.

4. The rain interference detection method based on a neural network according to claim 1, characterized in that, The rain detection neural network model is an MLP neural network model, and a spatial transformation network is further provided before the input layer of the MLP neural network model, which is used to extract features with spatial invariance from the input information as the input of the MLP neural network.

5. The rain interference detection method based on a neural network according to claim 1, wherein, the vehicle parameter information further includes: the current yaw rate of the vehicle, and / or the vehicle model; Before selecting the target point cloud that meets the detection threshold from the point cloud data based on a preset detection threshold, it further includes: retrieving the corresponding preset detection threshold according to the parameter information of the vehicle.

6. A rain interference detection device based on a neural network, wherein, it includes: an acquisition module, configured to acquire vehicle parameter information and point cloud data scanned by a radar; a screening module, configured to select a target point cloud that meets the detection threshold from the point cloud data based on a preset detection threshold; the detection threshold includes a spatial position area; an input module, configured to input the selected target point cloud and the vehicle parameter information into a trained rain detection neural network model; an identification module, configured to identify and output whether each target point cloud is a rain interference point cloud through the rain detection neural network model; the detection threshold further includes: a Doppler velocity interval and a reflection energy interval; the screening module specifically includes: a position screening sub-module, a velocity screening sub-module, and an energy screening sub-module; wherein: the position screening sub-module is configured to select the point cloud that falls into the spatial position area from the point cloud data as the first point cloud; the velocity screening sub-module is configured to screen the second point cloud that is within a preset Doppler velocity interval from the first point cloud; the energy screening sub-module is configured to screen the point cloud that is within a preset reflection energy interval from the second point cloud as the target point cloud; wherein, the training of the rain detection neural network model includes: collecting point cloud data and vehicle parameter information of a rain interference scene; selecting, based on a preset spatial position area, the point cloud that falls into the preset spatial position area from the point cloud data of the rain interference scene as an initial training sample; identifying whether the scene corresponding to the initial training sample is a pure rain interference scene; when it is determined that the scene corresponding to the initial training sample is a pure rain interference scene, labeling each target point cloud in the initial training sample, and storing the labeled initial training sample as a positive sample together with the corresponding vehicle parameter information in a training sample set; training an initial neural network model through the training sample set until a rain detection neural network model that meets the requirements is obtained.

7. The rain interference detection device based on a neural network according to claim 6, wherein, the rain detection neural network model is an MLP neural network model, and a spatial transformation network is further provided before the input layer of the MLP neural network model, which is used to extract features with spatial invariance from the input information as the input of the MLP neural network.

Images