Wide dynamic pulse laser ranging method and system for neural network

Abstract

The application provides a wide dynamic pulse laser ranging method and system of a neural network, comprising the following steps: S1, collecting echo signals after a laser radar transmits laser signals; S2, extracting characteristic parameters of the transmitted and received waveforms; S3, detecting rising edges and falling edges of the characteristic parameters, and if both the rising edges and the falling edges are valid, the pulse detection is successful; S4, calculating a time difference of corresponding positions of a seed light waveform of the transmitted signals and a waveform of the echo signals; and S5, obtaining a distance between the laser radar and a target by a time-of-flight ranging method. The application has reduced operation amount, high measurement accuracy and strong adaptive capacity. The wide dynamic ranging can be realized by a single-channel ADC acquisition module. The system has few channels, a simple circuit structure and low power consumption. The trained neural network can accurately predict the ranging result according to the input waveform characteristics.

Description

Technical Field

[0001] This invention relates to lidar ranging methods, and in particular to a wide dynamic range pulsed laser ranging method and system based on a neural network. Background Technology

[0002] The basic principle of pulsed laser ranging is to actively emit a pulsed laser towards the target, then detect the laser echo at the illuminated point on the target, and measure the flight time of the emitted and received pulsed lasers to obtain the distance from the target to the lidar rangefinder. To achieve wide dynamic range ranging, such as from a few centimeters to several kilometers, pulsed laser ranging typically employs automatic gain control (AGC) or multi-channel amplification techniques. However, both methods have limitations. AGC uses the pulse amplitude from the previous moment to control and adjust the pulse amplitude at the next moment, which cannot overcome ranging failures caused by sudden changes in target distance. While multi-channel amplification can overcome ranging failures caused by sudden changes in target distance, it introduces technical problems such as channel bias, channel overlap, and channel gaps. Furthermore, increasing the number of channels complicates the ranging system hardware and increases power consumption. Both methods use Gaussian fitting to calculate the seed light waveform and the echo waveform, requiring a high signal-to-noise ratio for time-series discrimination, and the waveform must not saturate. Chinese patent document CN 110058254 A describes a method, system, and terminal for calculating the echo time of pulsed laser ranging. It describes a scheme for full waveform analysis and classification for identification; however, this scheme is similar to a fitting calculation scheme, with a large computational load and low efficiency when dealing with large-scale digital point cloud operations. Furthermore, it is difficult to perform full waveform analysis in the case of saturated waveforms. CN 110297223 A describes a method for detecting the number of pulsed laser echoes based on a BP neural network, but it lacks specific algorithms and parameters, making it difficult to implement. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a wide dynamic range pulse laser ranging method and system based on neural networks, which can reduce the amount of computation and improve the accuracy of ranging.

[0004] Another technical problem to be solved by the present invention is to provide a neural network wide dynamic range pulsed laser ranging system, which uses a single-channel ADC acquisition module to acquire laser ranging with a wide dynamic range, thereby reducing the complexity of the system.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is: a wide dynamic range pulse laser ranging method based on neural networks, comprising the following steps: S1. After the lidar emits a laser signal, it collects the echo signal; S2. Extract the characteristic parameters of the transmitted and received waveforms; S3. Detect the rising and falling edges of the feature parameters. If both rising and falling edges are detected successfully, the pulse detection is successful. S4. Calculate the time difference between the corresponding positions of the seed light waveform of the transmitted signal and the echo signal waveform; S5. The distance between the lidar and the target is determined by the time-of-flight ranging method.

[0006] In the preferred embodiment, the waveform characteristics include the end time T of the rising edge. n The start time T of the falling edge m ; The value of AD on the rising edge is Y n-4 ~Y n The AD value at the falling edge is Y m ~Y m+4 .

[0007] In the preferred embodiment, in step S3, T n The AD sampling value corresponding to the sampling time is Y n ; The rising edge AD value of the waveform characteristics satisfies Y n-4 <Y n-3 <Y n-2 <Y n-1 <Y n and Y n ≥Y n+1 At that time, rising edge detection is valid; the falling edge AD value of the waveform feature satisfies Y. m >Y m+1 >Y m+2 >Y m+3 >Y m+4 and Y m-1 ≤Y m At that time, falling edge detection is valid; When both rising edge and falling edge detection are valid, the waveform characteristics match, and pulse detection is successful.

[0008] Preferably, in step S3, the number of times the waveform flattens out is P=mn, and P is 0 when no waveform flattens out.

[0009] Preferably, in step S3, the feature waveform detection is trained on a neural network training machine; The input for training is the waveform features of each ranging measurement, and the output is the actual distance between the lidar and the actual target corresponding to the waveform features of this ranging measurement. Multiple sets of training inputs and outputs constitute the training set. The neural network structure is a cascaded feedforward network. A sample in a cascaded feedforward network consists of p inputs and q outputs, with multiple hidden layers between the input layer and the output layer.

[0010] In the preferred scheme, a sample consists of 11 inputs, 2 outputs, and 2 hidden layers; The training process of a cascaded feedforward network consists of two sub-processes: the forward propagation sub-process of the working signal and the backward propagation sub-process of the error signal. In the forward propagation sub-process, let the weights between node i and node j be... The threshold for node j is The output value of each node is The output value of each node is determined by the output values ​​of all nodes in the previous layer, the weights of the current node relative to all nodes in the previous layer, the threshold of the current node, and the activation function. The activation function can be either a sigmoid function or a linear function. The specific calculation method is as follows: ;Formula 1 ;Formula 2 Assume all results of the output layer are The error function is as follows: ;Formula 2 The learning rule continuously adjusts the network's weights and thresholds along the fastest descent direction of the relative sum of squared errors. According to gradient descent, the correction of the weight vector is proportional to the change in the weights at the current position. The gradient of the j-th output node is: ;Formula 4 The activation function is: ;Formula 5 Differentiate the activation function: ;Formula 6 for: ;Formula 7 in: ;Formula 8 for: ;Formula 9 The above steps are used to calculate and adjust the weights between the hidden layer and the output layer, as well as the threshold of the output layer.

[0011] Weight adjustment amount; Learning rate; Local gradient; Activation function; In the preferred scheme, the threshold adjustment amounts for the input layer and hidden layer are calculated as follows: Assumption If is the weight between the k-th node in the input layer and the i-th node in the hidden layer, then: ;Formula 10 ;Formula 11 According to the gradient descent method, the weights and thresholds between the hidden layer and the output layer are adjusted as follows: ;Formula 12 The weights and thresholds between the input layer and the hidden layer are adjusted as follows: ;Formula 13 The seed light waveform parameters and return wave signal waveform parameters corresponding to each ranging process form a two-dimensional or one-dimensional vector data, which serves as the input to the neural network, and the corresponding actual distance serves as the output of the neural network.

[0012] In the preferred embodiment, the extracted waveform parameters and corresponding ranging results are input into the neural network training module, the network structure and training parameters are set, and the final parameters of the trained network are obtained. New wide dynamic range pulse waveforms are detected, and a neural network built using trained network parameters predicts new ranging results.

[0013] A system for the wide dynamic range pulsed laser ranging method of the above-mentioned neural network includes: a waveform acquisition module, a wide dynamic range pulse detection module, a waveform feature parameter extraction module, a waveform feature data transmission module, a neural network training module, a neural network parameter storage module, and a ranging prediction module. The waveform acquisition module is used to acquire the seed light waveform of the transmitted signal and the echo signal waveform. The waveform acquisition module is a single-channel ADC acquisition module. The wide dynamic range pulse detector module is used to detect the rising and falling edges of the seed light signal waveform and the echo signal waveform, respectively. The waveform feature parameter extraction module is used to extract the waveform features of the transmitted signal seed light waveform and the echo signal waveform; The waveform feature data transmission module is used to transmit waveform feature data; The neural network training module is used to input a certain number of extracted waveform features and corresponding ranging results into the neural network training module, set the network structure and training parameters for training, and obtain the final parameters of the trained network. The neural network parameter storage module is used to store neural network parameters; The ranging prediction module is used to input the waveform features of the acquired transmitted signal seed light waveform and echo signal waveform into a trained neural network to predict new ranging results.

[0014] In the preferred embodiment, the waveform features include the end time Tn of the rising edge and the start time Tm of the falling edge; The value of AD on the rising edge is Y n-i ~Y n The AD value at the falling edge is Y m ~Y m+j .

[0015] In the preferred scheme, T n The AD sampling value corresponding to the sampling time is Y n ; The rising edge AD value of the waveform characteristics satisfies Y n-a <...<Y n-2 <Y n-1 <Y n and Y n ≥Y n+1 At that time, rising edge detection is valid; the falling edge AD value of the waveform feature satisfies Y. m >Y m+1 >Y m+2 ...>Y m+b and Y m-1 ≤Y m At that time, falling edge detection is valid; a represents the number of samples on the rising edge, and b represents the number of samples on the falling edge; in this example, 4 to 5 samples are used.

[0016] When both rising edge and falling edge detection are valid, the waveform characteristics match, and pulse detection is successful.

[0017] This invention provides a wide dynamic range pulsed laser ranging method and system based on neural networks, which has the following advantages compared with the prior art: 1. Reduced computational load, high measurement accuracy, and strong adaptability.

[0018] 2. It can achieve wide dynamic range measurement with a single-channel ADC acquisition module.

[0019] 3. The system has a small number of channels, a simple circuit structure, and low power consumption.

[0020] 4. A well-trained neural network can accurately predict the ranging result based on the waveform characteristics of the input. Attached Figure Description

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of the neural network training process of the present invention.

[0022] Figure 2 This is a schematic diagram of waveform feature parameter extraction according to the present invention.

[0023] Figure 3 This is a schematic diagram illustrating the characteristic waveform detection performed by the wide dynamic range pulse detection module of the present invention. Detailed Implementation

[0024] Example 1: like Figure 1 As shown, a wide dynamic range pulsed laser ranging method using a neural network includes the following steps: S1. After the lidar emits a laser signal, it collects the echo signal; such as Figure 3 As shown, the echo signal differs from the transmitted seed wave signal to some extent. Moreover, due to the presence of interference waves, the waveform of the echo signal becomes asymmetrical, increasing the difficulty of identification.

[0025] S2. Extract the characteristic parameters of the transmitted and received waveforms; In the preferred embodiment, the waveform characteristics include the end time Tn of the rising edge and the start time T of the falling edge. m ; The value of AD on the rising edge is Y n-4 ~Y n The AD value at the falling edge is Y m ~Y m+4 .

[0026] S3. Detect the rising and falling edges of the feature parameters. If both rising and falling edges are detected successfully, the pulse detection is successful. In the preferred scheme, T n The AD sampling value corresponding to the sampling time is Y n ; The rising edge AD value of the waveform characteristics satisfies Y n-4 <Y n-3 <Y n-2 <Y n-1 <Y n and Y n ≥Y n+1 At that time, rising edge detection is valid; the falling edge AD value of the waveform feature satisfies Y. m >Y m+1 >Y m+2 >Y m+3 >Y m+4 and Y m-1 ≤Y m At that time, falling edge detection is valid; When both rising edge and falling edge detection are valid, the waveform characteristics match, and pulse detection is successful.

[0027] Preferably, in step S3, the number of times the waveform flattens out is P=mn, and P is 0 when no waveform flattens out.

[0028] S4. Calculate the time difference between the corresponding positions of the seed light waveform of the transmitted signal and the echo signal waveform; S5. The distance between the lidar and the target is determined by the time-of-flight ranging method.

[0029] Preferred, such as Figure 1 In step S3, the feature waveform detection is trained on the neural network training machine; The input for training is the waveform features of each ranging measurement, and the output is the actual distance between the lidar and the actual target corresponding to the waveform features of this ranging measurement. Multiple sets of training inputs and outputs constitute the training set. The neural network structure is a cascaded feedforward network. A sample in a cascaded feedforward network consists of p inputs and q outputs, with multiple hidden layers between the input layer and the output layer.

[0030] In the preferred scheme, a sample consists of 11 inputs, 2 outputs, and 2 hidden layers; The training process of a cascaded feedforward network consists of two sub-processes: the forward propagation sub-process of the working signal and the backward propagation sub-process of the error signal. In the forward propagation sub-process, let the weights between node i and node j be... The threshold for node j is The output value of each node is The output value of each node is determined by the output values ​​of all nodes in the previous layer, the weights of the current node relative to all nodes in the previous layer, the threshold of the current node, and the activation function. The activation function can be either a sigmoid function or a linear function. The specific calculation method is as follows: ;Formula 1 ;Formula 2 Assume all results of the output layer are The error function is as follows: ;Formula 2 The learning rule continuously adjusts the network's weights and thresholds along the fastest descent direction of the relative sum of squared errors. According to gradient descent, the correction of the weight vector is proportional to the change in the weights at the current position. The gradient of the j-th output node is: ;Formula 4 The activation function is: ;Formula 5 Differentiate the activation function: ;Formula 6 for: ;Formula 7 in: ;Formula 8 for: ;Formula 9 The above steps are used to calculate and adjust the weights between the hidden layer and the output layer, as well as the threshold of the output layer.

[0031] Weight adjustment amount; Learning rate; Local gradient; Activation function; In the preferred scheme, the threshold adjustment amounts for the input layer and hidden layer are calculated as follows: Assumption If is the weight between the k-th node in the input layer and the i-th node in the hidden layer, then: ;Formula 10 ;Formula 11 According to the gradient descent method, the weights and thresholds between the hidden layer and the output layer are adjusted as follows: ;Formula 12 The weights and thresholds between the input layer and the hidden layer are adjusted as follows: ;Formula 13 Each ranging process uses seed light waveform parameters and return wave signal waveform parameters to form a two-dimensional or one-dimensional vector data, which serves as the input to the neural network. The corresponding actual distance is used as the output of the neural network. During neural network training, the neural network structure, training termination conditions, and training and test samples are selected. The neural network training process in this patent is completed in MATLAB, and test samples are used to test the accuracy and generalization ability of the model.

[0032] In the preferred embodiment, the extracted waveform parameters and corresponding ranging results are input into the neural network training module, the network structure and training parameters are set, and the final parameters of the trained network are obtained. New wide dynamic range pulse waveforms are detected, and a neural network built using trained network parameters predicts new ranging results.

[0033] This invention employs a cascaded feedforward network for neural network training; multiple hidden layers can be used to fit nonlinear functions, and this invention uses two or more hidden layers to obtain ideal prediction results; according to the empirical formula: ;Formula 14 m and n are the number of neurons in the input and output layers, respectively, and a is a constant between [0, 10]. In this invention, the number of nodes in both hidden layers is set to 4 to satisfy an empirical formula. Training is performed using the backpropagation method for error signals, and the training process is as follows: Figure 1 As shown. Other training parameters are designed as follows: The minimum training error is set to 0.001, which is equivalent to 1 ps at a 1 GHz sampling rate. The learning rate is set to 0.05. Minimum performance gradient: 10^-20; The prediction module is implemented on the FPGA using HDL hardware language. The input is the waveform parameters output by the wide dynamic range detector module. Internally, it implements the neural network prediction calculation process and outputs the ranging result.

[0034] In this invention, the features of the seed wave and echo are not separated and are used as input for ranging training or prediction. The method of training a neural network using extracted waveform features significantly reduces computational load compared to the full waveform fitting analysis method, and also greatly improves ranging efficiency by using waveform features for ranging prediction.

[0035] Example 2: Based on Embodiment 1, a system for the wide dynamic range pulsed laser ranging method for the above-mentioned neural network includes: a waveform acquisition module, a wide dynamic range pulse detection module, a waveform feature parameter extraction module, a waveform feature data transmission module, a neural network training module, a neural network parameter storage module, and a ranging prediction module. The waveform acquisition module is used to acquire the seed light waveform of the transmitted signal and the echo signal waveform. The waveform acquisition module is a single-channel ADC acquisition module. The wide dynamic range pulse detector module is used to detect the rising and falling edges of the seed light signal waveform and the echo signal waveform, respectively. The waveform feature parameter extraction module is used to extract the waveform features of the transmitted signal seed light waveform and the echo signal waveform; The waveform feature data transmission module is used to transmit waveform feature data; The neural network training module is used to input a certain number of extracted waveform features and corresponding ranging results into the neural network training module, set the network structure and training parameters for training, and obtain the final parameters of the trained network. The neural network parameter storage module is used to store neural network parameters; The ranging prediction module is used to input the waveform features of the acquired transmitted signal seed light waveform and echo signal waveform into a trained neural network to predict new ranging results.

[0036] In the preferred embodiment, the waveform features include the end time Tn of the rising edge and the start time Tm of the falling edge; The value of AD on the rising edge is Y n-i ~Y n The AD value at the falling edge is Y m ~Y m+j .

[0037] In the preferred scheme, T n The AD sampling value corresponding to the sampling time is Y n ; The AD value of the rising edge of the waveform characteristic satisfies Y n-a <...<Y n-2 <Y n-1 <Y n and Y n ≥Y n+1 At that time, rising edge detection is valid; the AD value of the falling edge of the waveform feature satisfies Y. m >Y m+1 >Y m+2 ...>Y m+b Falling edge detection is valid when Ym-1≤Ym; a represents the number of samples on the rising edge, and b represents the number of samples on the falling edge; in this example, 4 to 5 samples are used.

[0038] When both rising edge and falling edge detection are valid, the waveform characteristics match, and pulse detection is successful.

[0039] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The embodiments and features described in these embodiments can be arbitrarily combined without conflict. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A wide dynamic range pulsed laser ranging method based on neural networks, characterized by: Includes the following steps: S1. After the lidar emits a laser signal, it collects the echo signal; S2. Extract the characteristic parameters of the transmitted and received waveforms; In step S2, the waveform features include the end time Tn of the rising edge and the start time Tm of the falling edge; The value of AD on the rising edge is Y n-4 ~Y n The AD value at the falling edge is Y m ~Y m+4 ; S3. Detect the rising and falling edges of the feature parameters. If both rising and falling edges are detected successfully, the pulse detection is successful. In step S3, T n The AD sampling value corresponding to the sampling time is Y n ; The sampling time of the waveform features satisfies Y n-4 <Y n-3 <Y n-2 <Y n-1 <Y n and Y n ≥Y n+1 At that time, rising edge detection is valid; the sampling time of the waveform features satisfies Y. m >Y m+1 >Y m+2 >Y m+3 >Y m+4 and Y m-1 ≤Y m At that time, falling edge detection is valid; When both rising edge and falling edge detection are valid, the waveform characteristics match and pulse detection is successful. In step S3, the feature waveform detection is trained on the neural network training machine; The input for training is the waveform features of each ranging measurement, and the output is the actual distance between the lidar and the actual target corresponding to the waveform features of this ranging measurement. Multiple sets of training inputs and outputs constitute the training set. The neural network structure is a cascaded feedforward network. A sample in a cascaded feedforward network consists of p inputs and q outputs, with multiple hidden layers between the input layer and the output layer. S4. Calculate the time difference between the corresponding positions of the seed light waveform of the transmitted signal and the echo signal waveform; S5. The distance between the lidar and the target is determined by the time-of-flight ranging method.

2. The wide dynamic range pulsed laser ranging method based on neural networks according to claim 1, characterized in that: In step S3, the number of times the waveform flattens is P=mn, and P is 0 when no waveform flattens.

3. The wide dynamic range pulsed laser ranging method based on a neural network according to claim 1, characterized in that: A sample consists of 11 inputs, 2 outputs, and 2 hidden layers; The training process of a cascaded feedforward network consists of two sub-processes: the forward propagation sub-process of the working signal and the backward propagation sub-process of the error signal. In the forward propagation sub-process, let the weights between node i and node j be... The threshold for node j is The output value of each node is The output value of each node is determined by the output values ​​of all nodes in the previous layer, the weights of the current node relative to all nodes in the previous layer, the threshold of the current node, and the activation function. The activation function can be either a sigmoid function or a linear function. The specific calculation method is as follows: Official (1); Official (2); Assume all results of the output layer are The error function is as follows: Official (3); The learning rule continuously adjusts the network's weights and thresholds along the fastest descent direction of the relative sum of squared errors. According to gradient descent, the correction of the weight vector is proportional to the change in the weights at the current position. The gradient of the j-th output node is: Official (4); The activation function is: Official (5); Differentiate the activation function: Official (6); for: Official (7); in: Official (8); Corresponding threshold for: Official (9); The above steps are used to calculate and adjust the weights between the hidden layer and the output layer, as well as the threshold of the output layer. Weight adjustment amount; Learning rate; Local gradient; Activation function.

4. The wide dynamic range pulsed laser ranging method based on a neural network according to claim 3, characterized in that: The threshold adjustment amounts for the input layer and hidden layer are calculated as follows: Assumption If is the weight between the k-th node in the input layer and the i-th node in the hidden layer, then: Official (10); Official (11); According to the gradient descent method, the weights and thresholds between the hidden layer and the output layer are adjusted as follows: Official (12); The weights and thresholds between the input layer and the hidden layer are adjusted as follows: Official (13); The seed light waveform parameters and return wave signal waveform parameters corresponding to each ranging process form a two-dimensional or one-dimensional vector data, which serves as the input to the neural network, and the corresponding actual distance serves as the output of the neural network.

5. The wide dynamic range pulsed laser ranging method based on a neural network according to claim 1, characterized in that: The extracted waveform parameters and corresponding ranging results are input into the neural network training module. The network structure and training parameters are set for training to obtain the final parameters of the trained network. New wide dynamic range pulse waveforms are detected, and a neural network built using trained network parameters predicts new ranging results.

6. A system for a wide dynamic range pulsed laser ranging method using a neural network according to any one of claims 1 to 5, characterized in that: It includes: a waveform acquisition module, a wide dynamic range pulse detection module, a waveform feature parameter extraction module, a waveform feature data transmission module, a neural network training module, a neural network parameter storage module, and a ranging prediction module; The waveform acquisition module is used to acquire the seed light waveform of the transmitted signal and the echo signal waveform. The waveform acquisition module is a single-channel ADC acquisition module. The wide dynamic range pulse detector module is used to detect the rising and falling edges of the seed light signal waveform and the echo signal waveform, respectively. The waveform feature parameter extraction module is used to extract the waveform features of the transmitted signal seed light waveform and the echo signal waveform; The waveform characteristics include the end time Tn of the rising edge and the start time Tm of the falling edge; the AD value of the rising edge is Y. n-i ~Y n The AD value at the falling edge is Y m ~Y m+j ; T n The AD sampling value corresponding to the sampling time is Y n ; The rising edge AD value of the waveform characteristics satisfies Y n-a <...<Y n-2 <Y n-1 <Y n and Y n ≥Y n+1 At that time, rising edge detection is valid; the falling edge AD value of the waveform feature satisfies Y. m >Y m+1 >Y m+2 >Y m+3 >Y m+4 and Y m-1 ≤Y m At that time, falling edge detection is valid; a represents the number of samples on the rising edge, and b represents the number of samples on the falling edge; When both rising edge and falling edge detection are valid, the waveform characteristics match and pulse detection is successful. The waveform feature data transmission module is used to transmit waveform feature data; The neural network training module is used to input a certain number of extracted waveform features and corresponding ranging results into the neural network training module, set the network structure and training parameters for training, and obtain the final parameters of the trained network. The neural network parameter storage module is used to store neural network parameters; The ranging prediction module is used to input the waveform features of the acquired transmitted signal seed light waveform and echo signal waveform into a trained neural network to predict new ranging results.

Images