A millimeter wave radar weather recognition method based on neural network

Abstract

The present invention relates to a millimeter wave radar weather recognition method based on neural network. The radar emits electromagnetic waves forward, which are reflected back by rain and snow targets. The electromagnetic waves with front environment information received by the radar receiving antenna are analyzed, and the above data are input into a preset weather state recognition model for weather classification and recognition to obtain current weather state information. The weather state recognition model includes a convolutional neural network and a gated recurrent neural network. The radar data is classified and recognized for weather state through the above network to obtain a target classification label, i.e., current weather state information. According to the current state information, the radar performs corresponding clutter filtering. The method only needs a single millimeter wave radar for detection and recognition, and does not need to add other sensors to assist in weather recognition, thereby avoiding the complexity of the system and enhancing the adaptability and stability of the radar.

Description

Technical Field

[0001] The present invention relates to the technical field of millimeter wave radar, and in particular to a millimeter wave radar weather recognition method based on neural network. Background Art

[0002] As the number of motor vehicles owned by the public increases year by year, the intelligent parking system has developed rapidly. The deployment of gate equipment, accurate perception of vehicles entering and exiting, and smooth and unimpeded safe entry into the parking lot have put forward higher requirements for vehicle monitoring sensors; millimeter-wave radar has high resolution and can accurately detect the position, speed and direction of movement of vehicles, thereby improving the accuracy of the parking lot gate system, and millimeter-wave radar can detect longer distances, suitable for parking lots of various sizes, and can respond promptly to vehicles entering and exiting quickly; millimeter-wave radar can detect multiple targets at the same time, and can effectively manage vehicle entry and exit and avoid congestion in parking lots with large traffic volume; compared with other sensors, millimeter-wave radar has lower maintenance requirements, relatively low failure rate, and lower maintenance cost; due to the above characteristics, millimeter-wave radar has gradually become an important sensor for intelligent parking systems.

[0003] However, most parking barriers are currently located outdoors, and the weather in outdoor scenes is an important factor affecting the performance of the barrier millimeter-wave radar. In rainy and snowy weather, the electromagnetic waves emitted by the barrier radar are reflected back by rain and snow targets, causing the radar to detect rain and snow clutter targets, thereby causing radar false alarms and reducing the recognition accuracy of the barrier radar. Therefore, millimeter-wave radar is needed to identify weather conditions and provide information for the subsequent use of different clutter suppression parameters for different weather conditions. The existing technology usually performs weather recognition operations by assigning values ​​and weighting operations on conditional thresholds through experience, but the weather is constantly changing and unpredictable, which makes the fixed threshold too one-sided and not universal. Summary of the invention

[0004] In view of this, the present invention provides a millimeter-wave radar weather recognition method based on a neural network, which effectively solves the problem that the millimeter-wave radar in the prior art uses a fixed threshold to judge the weather state and is not universal. By inputting the environmental feature data collected by the radar, it can accurately identify the conventional weather conditions such as light rain, moderate rain, light snow, moderate snow, etc. and allow the radar to make adjustments according to different weather conditions.

[0005] To achieve the above object, the present invention provides a millimeter wave radar weather recognition method based on neural network, comprising the following steps:

[0006] S1. Get the original ADC radar signal of the current weather;

[0007] S2, preprocessing the original radar data to obtain a range-Doppler heat map;

[0008] S3, inputting the preprocessed data into a weather recognition model for weather classification and recognition;

[0009] S301, splitting the sequence of the range-Doppler heat map in chronological order, and respectively inputting the range-Doppler heat map corresponding to each moment into a convolutional neural network CNN to extract an 8-channel feature vector sequence;

[0010] S302, sequentially convert the 1-channel feature vector corresponding to each moment into Input into the gated recurrent neural network GRU to get the output sequence ,in ;

[0011] S303: Output the sequence Input to the classifier network, split into multi-dimensional vectors through the split multi-dimensional vector network and enter the fully connected layer, and obtain the target distance-Doppler heat map data through the linear rectification function ReLU;

[0012] S304, inputting the target distance-Doppler heat map data into a preset weather state recognition model, and outputting a weather classification label corresponding to each distance-Doppler heat map;

[0013] S4. Output current weather status information and adaptively adjust radar parameters.

[0014] Preferably, the millimeter wave radar transmits electromagnetic waves through the transmitting antenna and receives the reflected echo data through the connecting antenna within one detection cycle. The mixer combines the transmitting signal and the receiving signal together to generate an intermediate frequency signal. The ADC module samples the intermediate frequency signal to obtain the original ADC radar signal of the current weather.

[0015] Preferably, a data pre-whitening method is used for data pre-processing, and range FFT and Doppler FFT are used to obtain a radar range-Doppler heat map.

[0016] Preferably, the convolutional neural network CNN comprises a two-layer structure, the first layer contains 4 channels The convolution layer and the convolution kernel Maximum pooling layer, the second layer contains 8 channels The convolution layer and the convolution kernel Max pooling layer.

[0017] Preferably, the gated recurrent neural network GRU includes a single hidden layer, and the hidden layer dimension is 208.

[0018] Preferably, the split multidimensional vector network includes a convolution layer with 3 channels and 33 convolution kernels and a 22 maximum pooling layer; the linear rectification function ReLU calculation method is:

[0019] ;

[0020] The dropout ratio in the fully connected layer is set to 0.5.

[0021] Preferably, the radar performing adaptive adjustment in step S4 includes:

[0022] When the current weather is clear, the weather status flag is 0. After being transmitted to the radar, the CFAR parameters use normal parameters to effectively detect the target;

[0023] When the current weather is rainy, the weather status flag is 1. After being transmitted to the radar, the threshold factor threshold in the CFAR parameter is increased by 1dB to effectively detect the target;

[0024] When the current weather is snowing, the weather status flag is 2. After being transmitted to the radar, the threshold factor threshold in the CFAR parameter is increased by 2dB to effectively detect the target.

[0025] Preferably, after CFAR detects the point cloud, the point cloud data result is filtered twice by setting a threshold on the area in front of the radar.

[0026] When the weather status flag is 0, the regional threshold parameters are fixed;

[0027] When the weather status flag is 1, the energy value threshold of each area is +50;

[0028] When the weather status flag is 2, the energy value threshold of each area is +100.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present invention detects and identifies the surrounding environment through a single millimeter-wave radar, and can complete weather identification without adding other sensors to assist, thus avoiding the complexity of the system; 2D-FFT is used to obtain the distance-Doppler heat map and use it as the neural network input, without the need for other processing and data features, and the high sensitivity can quickly and accurately capture environmental information with low data complexity, thereby improving the calculation efficiency;

[0031] The present invention utilizes the efficient feature extraction capability of CNN to complete the feature extraction of Doppler heat maps, utilizes the temporal correlation characterization function of GRU on time series data to complete the temporal correlation description of feature map sequences, and introduces convolutional structures to extract multi-scale combined features, assist high-level feature learning, and enrich the feature input of GRU. However, it does not require complex feature extraction and processing of input data, and only needs to obtain data feature information through different network operations to achieve high-precision weather recognition.

[0032] The network model carried by the present invention can detect and identify the changes in the current weather in real time after data training, and adjust the millimeter-wave radar parameters in real time according to the current weather, thereby improving the adaptability and stability of the radar under the weather. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] FIG1 is a flow chart of the present invention based on millimeter wave radar weather identification;

[0034] Figure 2 This is a GRU network unit diagram of the present invention;

[0035] Figure 3 This is a diagram of the data processing and algorithm structure of the present invention;

[0036] Figure 4 is the rain and fog attenuation curve in the embodiment of the present invention;

[0037] Figure 4 In the figure, curve a is light rain (rainfall 0.25 mm / h); curve b is light rain (rainfall 1 mm / h); curve c is heavy rain (rainfall 4 mm / h); curve d is torrential rain (rainfall 16 mm / h); and curve e is fog, with a visibility of 600 m (water content 0.032 g / m 3 ); f is fog, its concentration is visibility 120m (water content 0.32g / m 3 ); g is dense fog, with a visibility of 30m (water content 2.3g / m 3 ). DETAILED DESCRIPTION

[0038] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0039] In order to effectively improve the recognition accuracy of weather such as rain and snow and eliminate the influence of weather factors on radar target recognition, the present invention provides a millimeter wave radar weather recognition method based on a neural network, comprising the following steps:

[0040] S1. In a detection cycle, the radar transmitting antenna transmits electromagnetic waves forward, which are reflected back when encountering rain or snow targets. The mixer combines the transmitting signal and the receiving signal to generate an intermediate frequency signal. The ADC module samples the intermediate frequency signal to obtain the original ADC radar signal of the current weather.

[0041] S2, using the data pre-whitening method to pre-process the electromagnetic waves with the front environment information received by the radar receiving antenna, so as to eliminate data redundancy, and using the range-FFT and Doppler-FFT processing to obtain the range-Doppler heat map (RD map);

[0042] S3, input the preprocessed data into the weather recognition model for weather classification and recognition; the overall network structure is mainly divided into three parts: CNN, GRU and classifier;

[0043] S301, splitting the sequence of range-Doppler heat maps in chronological order, and respectively inputting the range-Doppler heat maps corresponding to each moment into a convolutional neural network CNN to extract an 8-channel feature vector sequence;

[0044] Convolutional neural network CNN consists of two layers, the first layer contains 4 channels The convolution layer and the convolution kernel Maximum pooling layer, the second layer contains 8 channels The convolution layer and the convolution kernel Maximum pooling layer; CNN network has the following advantages: it can use the original image as input directly, avoiding a lot of preprocessing; local receptive field acquisition features can reduce the impact of image translation, scaling and rotation; parameter sharing mechanism can reduce the complexity of network structure and improve computational efficiency; downsampling makes full use of local correlation information, which can reduce data processing and retain structural information; the abstract expression of Doppler heat map can be effectively extracted through nonlinear structure, while reducing manual intervention;

[0045] The basis for selecting the number of CNN network layers, convolution kernel size and number of channels is as follows: referring to similar image feature extraction tasks, such as image classification and target detection, CNN network structure design, selecting more common network layer combination ideas and structural parameter settings; then, based on the characteristics of the range-Doppler heat map, ensuring that the network will not overfit the training set during training and maintain generalization to the test set, select the appropriate number of convolution kernels and convolution layers; after inputting data, it can automatically complete feature extraction and learning time series patterns, and the CNN network will finally directly output the detection results;

[0046] S302, sequentially convert the 1-channel feature vector corresponding to each moment into Input into the gated recurrent neural network GRU to get the output sequence ,in ;

[0047] The gated recurrent neural network GRU includes a single hidden layer with a hidden layer dimension of 208. The single-layer GRU is used to speed up the convergence speed and improve the computing efficiency of the network. The GRU network is a type of recurrent neural network. Like LSTM (Long-Short Term Memory), it is also proposed to solve problems such as long-term memory and gradient in back propagation. In addition, GRU has fewer parameters, faster convergence speed, simple structure, fast iteration and easier training.

[0048] The GRU network structure is as follows Figure 2 As shown,

[0049] ;

[0050] ;

[0051] ;

[0052] ;

[0053] The GRU network has only two gates, corresponding to the output update gate vector and resetgate vector ; The update gate is responsible for controlling the status information of the previous moment The impact on the current state, the larger the value of the update gate is, the better the state information of the previous moment is. The more you bring in, the more control the reset gate takes;

[0054] S303, output sequence Input to the classifier network, including split multidimensional vector network, fully connected layer and linear rectification function ReLU. Split the multidimensional vector network into multidimensional vectors and enter the fully connected layer. The target distance-Doppler heat map data is obtained through the linear rectification function ReLU.

[0055] The split multidimensional vector network includes a 3-channel 33-convolution kernel convolution layer and a 22-maximum pooling layer; the linear rectification function ReLU calculation method is:

[0056] ;

[0057] The dropout ratio in the fully connected layer is set to 0.5;

[0058] S304, inputting the target distance-Doppler heat map data into a preset weather state recognition model, and outputting a weather classification label corresponding to each distance-Doppler heat map;

[0059] S4, output current weather status information and adaptively adjust radar parameters;

[0060] Adaptive adjustments include:

[0061] When the current weather is clear, the weather status flag is 0. After being transmitted to the radar, the CFAR parameters use normal parameters to effectively detect the target;

[0062] When the current weather is rainy, the weather status flag is 1. After being transmitted to the radar, the threshold factor threshold in the CFAR parameter is increased by 1dB to effectively detect the target;

[0063] When the current weather is snowy, the weather status flag is 2. After being transmitted to the radar, the threshold factor threshold in the CFAR parameter is increased by 2dB to effectively detect the target;

[0064] When CFAR detects the point cloud, it performs secondary filtering on the point cloud data results by setting a threshold limit on the area in front of the radar.

[0065] When the weather status flag is 0, the regional threshold parameters are fixed;

[0066] When the weather status flag is 1, the energy value threshold of each area is +50;

[0067] When the weather status flag is 2, the energy value threshold of each area is +100.

[0068] In order to maintain the detection performance of the millimeter-wave radar, through multiple verifications, the CNN+GRU network structure not only has a good accuracy rate, but also has a very low processing delay, which meets the requirements of timely output of detection results, facilitates the conversion of radar weather modes, and thus reduces the misjudgment of the millimeter-wave radar caused by weather.

[0069] The scheme of the present invention is aimed at recognition in normal weather, that is, light rain, moderate rain, light snow, moderate snow and other common weather conditions; drizzle has little effect on the performance of millimeter wave radar; and extreme weather conditions are not within the scope of discussion of the scheme of the present invention because data is difficult to collect and reproduce;

[0070] like Figure 4 The attenuation rate of the millimeter-wave radar beam for rain is small, but the background noise is still different from that on a sunny day. Therefore, increasing the CFAR threshold by 1dB can avoid this slight influence and raise the background noise, leading to false detection. In snowy weather, snowfall will produce a high attenuation rate, and the increase in background noise caused by signal attenuation and weakened echo strength will have a greater impact on the detection target. Therefore, the CFAR threshold needs to be increased by 2dB to avoid the influence of background noise when detecting targets.

[0071] For extreme weather, more real data is needed to dynamically adjust and optimize the thresholds.

[0072] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A millimeter wave radar weather recognition method based on neural network, characterized in that: The following steps are involved: S1. Get the original ADC radar signal of the current weather; S2, preprocessing the original radar data to obtain a range-Doppler heat map; S3, inputting the preprocessed data into a weather recognition model for weather classification and recognition; S301, splitting the sequence of the range-Doppler heat map in chronological order, and respectively inputting the range-Doppler heat map corresponding to each moment into a convolutional neural network CNN to extract an 8-channel feature vector sequence; S302, sequentially convert the 1-channel feature vector corresponding to each moment into Input into the gated recurrent neural network GRU to get the output sequence ,in ; S303: Output the sequence Input to the classifier network, split into multi-dimensional vectors through the split multi-dimensional vector network and enter the fully connected layer, and obtain the target distance-Doppler heat map data through the linear rectification function ReLU; S304, inputting the target range-Doppler heat map data into a preset weather state recognition model, and outputting a weather classification label corresponding to each range-Doppler heat map; S4. Output current weather status information and adaptively adjust radar parameters.

2. The neural network-based millimeter wave radar weather recognition method according to claim 1, characterized in that: During a detection cycle, the millimeter-wave radar transmits electromagnetic waves through the transmitting antenna and receives the reflected echo data through the connecting antenna. The mixer combines the transmitting signal and the receiving signal to generate an intermediate frequency signal. The ADC module samples the intermediate frequency signal to obtain the original ADC radar signal of the current weather.

3. The neural network-based millimeter wave radar weather recognition method according to claim 1, characterized in that: The data pre-whitening method is used for data preprocessing, and the radar range-Doppler heat map is obtained by using range FFT and Doppler FFT processing.

4. The neural network-based millimeter wave radar weather recognition method according to claim 1, characterized in that: The convolutional neural network CNN includes a two-layer structure, the first layer contains 4 channels The convolution layer and the convolution kernel Maximum pooling layer, the second layer contains 8 channels The convolution layer and the convolution kernel Max pooling layer.

5. The neural network-based millimeter wave radar weather recognition method according to claim 1, characterized in that: The gated recurrent neural network GRU includes a single hidden layer, and the hidden layer dimension is 208.

6. The neural network-based millimeter wave radar weather recognition method according to claim 1, characterized in that: The split multi-dimensional vector network includes a convolution layer with 3 channels and 33 convolution kernels and a 22 maximum pooling layer; the linear rectification function ReLU calculation method is: ； The dropout ratio in the fully connected layer is set to 0.

5.

7. The neural network-based millimeter wave radar weather recognition method according to claim 1, characterized in that: In step S4, the radar performs adaptive adjustment including: When the current weather is clear, the weather status flag is 0. After being transmitted to the radar, the CFAR parameters use normal parameters to effectively detect the target; When the current weather is rainy, the weather status flag is 1. After being transmitted to the radar, the threshold factor threshold in the CFAR parameter is increased by 1dB to effectively detect the target; When the current weather is snowing, the weather status flag is 2. After being transmitted to the radar, the threshold factor threshold in the CFAR parameter is increased by 2dB to effectively detect the target.

8. The neural network-based millimeter wave radar weather recognition method according to claim 7, characterized in that: When CFAR detects the point cloud, it performs secondary filtering on the point cloud data results by setting a threshold limit on the area in front of the radar. When the weather status flag is 0, the regional threshold parameters are fixed; When the weather status flag is 1, the energy value threshold of each area is +50; When the weather status flag is 2, the energy value threshold of each area is +100.

Images