Radar information processing parameter automatic generation method based on perception prediction
Through the method based on perception prediction, radar information processing parameters are automatically generated, which solves the adaptive problem of reinforcement learning algorithms in complex open scenarios, and improves radar detection performance and scene adaptability.
Patent Information
- Application Number
- CN202411481257.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-23
- Publication Date
- 2025-02-14
AI Technical Summary
In complex open scenarios, reinforcement learning algorithms are difficult to effectively use, resulting in the inability to adapt to radar information processing parameters and affecting detection performance.
The radar information processing parameter automatic generation method based on perception prediction is adopted, and future information processing parameters are automatically given through area division, dimensionality reduction coding and timing analysis and prediction, combined with multi-party information.
It effectively enhances the scenario adaptability of radar information processing, fully utilizes the resource potential of radar system, and reduces the workload of operators.
Smart Images
Figure CN119439097A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of radar information processing and is a method for realizing radar adaptive parameter generation. Background Art
[0002] In different detection scenarios, different radar information processing parameters will have different effects on radar detection performance. How to automatically adjust radar information processing parameters according to the external environment is one of the key issues that relevant researchers are concerned about. In recent years, as a decision-making algorithm, reinforcement learning has received increasing attention in many fields, and has demonstrated capabilities comparable to or even exceeding those of humans in areas such as autonomous driving and chess and card games. However, although most reinforcement learning algorithms can perform well in closed scenarios such as games and chess and card games, their performance will deteriorate in complex open scenarios without clear boundaries, and they are prone to problems such as overfitting. How to apply the decision-making ability of reinforcement learning in open scenarios faced by radar is a problem of both scientific and engineering significance. Summary of the invention
[0003] The present invention proposes a method for automatically generating radar information processing parameters based on perception prediction, which realizes the automatic generation of radar information processing parameters and can effectively enhance the scene adaptability of radar information processing.
[0004] The present invention first divides the external environment into regions according to the echo information received by the radar, and the radar data and information processing parameters of the current cycle in each region are input into a decision model composed of an encoder, a predictor and a decision maker, and the model gives the information processing parameters of the region in the next cycle. The implementation steps of the present invention are:
[0005] Step 1: After the radar receives the echo, the image composed of the radar echo is sent to the algorithm that can realize region division. According to the prior, the external environment is divided into multiple regions;
[0006] Step 2: In each area, the original radar echo data of the current period is spliced with the intermediate result or final result of the information processing under the current information processing parameters;
[0007] Step 3: The concatenated result of each region is fed into the encoder of the decision model, which compresses the high-dimensional, information-sparse radar data of the current period into a low-dimensional, information-dense feature vector;
[0008] Step 4: The predictor infers the feature vector corresponding to the next radar data cycle based on the feature vector corresponding to the current period radar data, the feature vector at the historical moment, and the current information processing parameter information;
[0009] Step 5: The decision maker gives the information processing parameters for the specific area of the next cycle based on the feature vector obtained by the encoder, the feature vector output by the predictor, and the information processing parameters of the current cycle.
[0010] Furthermore, the radar information processing parameters include various algorithm adjustable parameters in matched filtering, Doppler processing, detection, tracking, and identification.
[0011] Furthermore, the decision maker is any reinforcement learning algorithm that can achieve continuous control of parameters.
[0012] The beneficial effects of the present invention include:
[0013] In response to the adaptive needs of radar information processing parameters, the complex external environment of the radar is first divided into several small areas through regional division, and then the echoes in each area are dimensionality reduced and encoded, and time series analysis and prediction are performed in the low-dimensional feature space. Combining information from multiple parties, future information processing parameters are automatically given, which can effectively enhance the scene adaptability of radar information processing, give full play to the resource potential of the radar system itself, and reduce the workload of operators.
[0014] The present invention is described in further detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 The figure is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION
[0016] The present invention will be further described below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited by the implementing regulations.
[0017] The present invention proposes a method for automatically generating radar information processing parameters based on perception prediction. First, the external environment is divided into regions according to the echo information received by the radar. The radar data and the information processing parameters of the current cycle in each region are input into a decision model composed of an encoder, a predictor and a decision maker. The model gives the information processing parameters of the region in the next cycle. The specific steps are as follows:
[0018] (1) After receiving the echo, the radar sends the image composed of the radar echo to an algorithm that can realize region division. According to the prior, the external environment is divided into multiple regions. In this embodiment, the U-Net network widely used in the field of image segmentation is used to divide the radar data into a clutter region and a noise region. For each region, the image is scaled to a fixed size, which is 512*512 in this embodiment.
[0019] (2) In each region, the radar raw echo data of the current period is spliced with the intermediate result or final result of the information processing under the current information processing parameters. In this embodiment, the parameters of the information processing include the CFAR detection threshold and the tracking start parameter.
[0020] (3) The result of splicing each region is sent to the encoder of the decision model, and the encoder compresses the high-dimensional, information-sparse radar data of the current period into a low-dimensional, information-dense feature vector. In this embodiment, the encoder adopts an encoder module of a variational autoencoder based on a convolutional neural network, including five layers of convolution, and the number of channels of each convolution layer is 4, 8, 16, 32, and 64, respectively, and the ReLU commonly used in neural networks is used as the activation function. The encoder first compresses the 512*512 radar data of each region into a 16*16*64 feature map, and then maps the feature map to a feature vector of length 128 through a fully connected layer. The feature vector is the representation of the current data in a specific area.
[0021] (4) The predictor infers the feature vector corresponding to the next round of radar data based on the feature vector corresponding to the current period radar data, the feature vector at the historical moment, and the current information processing parameter information. In this embodiment, the Transformer decoder is used as the sequence prediction model, and the model predicts the feature vector of the next round of radar data based on the feature vectors corresponding to the previous five rounds of radar data.
[0022] (5) The decision maker gives the information processing parameters of the specific area of the next cycle according to the feature vector obtained by the encoder, the feature vector output by the predictor, and the information processing parameters of the current cycle. This embodiment adopts the classic Deterministic Policy Gradient (DPG) algorithm in reinforcement learning to realize the continuous control of the CFAR detection threshold and the tracking start parameter. In the training stage, the built-in value network and policy network can be trained by the time difference algorithm and the gradient ascent algorithm respectively; in the inference stage, the feature vector of the current data in the specific area, the feature vector output by the predictor, and the CFAR detection threshold and tracking start parameter of the current cycle are used as input, and the output is the CFAR detection threshold and tracking start parameter of the next cycle.
Claims
1. A method for automatically generating radar information processing parameters based on perception prediction, characterized in that: Step 1: After the radar receives the echo, the image composed of the radar echo is sent to the algorithm that can realize region division. According to the prior, the external environment is divided into multiple regions; Step 2: In each area, the original radar echo data of the current period is spliced with the intermediate result or final result of the information processing under the current information processing parameters; Step 3: The concatenated result of each region is fed into the encoder of the decision model, which compresses the high-dimensional, information-sparse radar data of the current period into a low-dimensional, information-dense feature vector; Step 4: The predictor infers the feature vector corresponding to the next radar data cycle based on the feature vector corresponding to the current period radar data, the feature vector at the historical moment, and the current information processing parameter information; Step 5: The decision maker gives the information processing parameters for the specific area of the next cycle based on the feature vector obtained by the encoder, the feature vector output by the predictor, and the information processing parameters of the current cycle.
2. The method for automatically generating radar information processing parameters based on perception prediction according to claim 1, characterized in that: The radar information processing parameters include various algorithm adjustable parameters in matched filtering, Doppler processing, detection, tracking and identification.
3. The method for automatically generating radar information processing parameters based on perception prediction according to claim 1, characterized in that: The decision maker is any reinforcement learning algorithm that can achieve continuous control of parameters.