Lidar pose error online measurement and calibration method based on neural network

Abstract

The application relates to a neural network-based laser radar attitude error online measurement and calibration method in the technical field of laser radar motors. T The neural network-based laser radar attitude error online measurement and calibration method comprises the following steps of inputting historical attitude angle data w T into a lightweight neural network to predict a scale factor A ‑1 and a nonlinear time-varying error value e k , then obtaining real-time errors and calibration values based on A ‑1 , e k , laser radar motor k attitude parameters w k at each moment. The application realizes error compensation and noise suppression by learning the nonlinear mapping of a time sequence, does not need to construct a complex analytical physical model, the lightweight neural network can directly learn error correction amounts, avoids displaying and decomposing various error components, has strong generalization ability, is suitable for real-time application and a resource-restricted embedded environment, and realizes laser radar motor system attitude parameter error online measurement and calibration.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of laser radar motor, in particular to a laser radar attitude error online measurement and calibration method based on a lightweight neural network, a laser radar motor and a computer readable storage medium. BACKGROUND

[0002] In the field of meteorological monitoring, the attitude parameters of the laser radar motor system are measured by an angle encoder in the motor system. In the motor system, the motor shaft drives the rotation of the rotating mechanical structure. The angle encoder provides key carrier attitude information by reading the number of rotations of the motor shaft, thereby realizing the scanning and positioning of the laser radar system in the horizontal direction / tilt direction. However, in the motor servo system, the angle feedback signal is easily affected by sensor accuracy, structural machining accuracy, mechanical assembly and transmission error, signal transmission delay, environmental factors and the like under actual operating conditions. The measurement error of the angle encoder is manifested as deterministic error and random error, which causes deviations in the positioning accuracy and stability of the laser radar signal, and non-expected instantaneous displacement, deformation or flicker of the target in the pixel coordinates between adjacent frames, thereby affecting the measurement stability. Deterministic error mainly includes zero drift, scale factor error, axis coaxiality error, axial offset error, system nonlinear error, etc. Random error mainly includes electronic noise, environmental disturbance, etc. Deterministic error is based on physical model and prior knowledge, and is calibrated through system calibration, error modeling and compensation algorithm. However, the actual system has nonlinear time-varying, multi-source coupling and high-dimensional characteristic error, and has limited real-time performance and adaptability. Random error is based on statistical hypothesis, and is processed by using a mean filter, a Kalman filter and the like designed by a person to reduce noise. The performance of the random error is reduced when facing non-Gaussian noise. Therefore, the above-mentioned methods for measuring and calibrating the attitude parameter error of the laser radar motor system have limited real-time performance and poor generalization ability, and it is difficult to effectively compensate. SUMMARY

[0003] In order to solve the technical problem that the existing technology has poor generalization ability and is difficult to effectively compensate by constructing a complex physical model to correct the attitude parameter error of the laser radar motor system, the present application provides a laser radar attitude error online measurement and calibration method based on a neural network.

[0004] In the first aspect, the present application provides a laser radar attitude error online measurement and calibration method based on a neural network, which inputs historical attitude angle data T of a length of w T into a lightweight neural network to predict the scale factor A -1 and the nonlinear time-varying error value e k , and then calibrates the attitude angle data based on A-1 , e k LiDAR motor k Attitude angle data at time step w k Obtain real-time error and the recalibrated attitude angle data .

[0005] The lightweight neural network includes: a scaling factor learning module, an error estimation module, and an adaptive fusion module. The scaling factor learning module is used to... w T After performing global feature extraction and nonlinear feature transformation, the following is obtained: A -1 and will A -1 and w k The scaling correction value is obtained by performing multiplication. rec The error estimation module includes: a temporal learning layer, a multi-head attention mechanism layer, and an error estimation layer; the temporal learning layer is used to extract errors sequentially over time. w k Temporal characteristics tf The multi-head attention mechanism layer is used to extract information through a multi-head attention mechanism. tf Enhanced features sf The error estimation layer is used to... sf Nonlinear mapping is e k The adaptive fusion module includes: stitching layer one, multilayer perceptron one, and stitching layer two. A pair of stitching layers... rec and e k The fused feature vector is obtained by concatenating the features. mf Multilayer perceptron - used for learning mf To correct the mapping; the splicing layer two combines the output of the multilayer perceptron one with... rec Perform addition to obtain .

[0006] Secondly, a neural network-based online measurement and calibration system for lidar attitude error is proposed, which utilizes the neural network-based online measurement and calibration method for lidar attitude error described in the first aspect. The online lidar attitude error measurement and calibration system includes: an acquisition module and a correction module. The acquisition module is used to acquire data from the lidar motor. k Attitude angle data at time step w k The correction module is used to process historical attitude angle data of length T. w T Input into the calibration neural network to predict the scaling factor A-1 and nonlinear time-varying error values e k , and then based on A -1 , e k , a laser radar motor k attitude angle data at the moment w k get real-time error and the attitude angle data after recalibration .

[0007] The third aspect also proposes a laser radar motor based on a neural network, which is controlled by the attitude angle data processed by the laser radar attitude error online measurement and calibration method based on a neural network in the first aspect.

[0008] The fourth aspect also proposes a computer readable storage medium, which stores computer instructions, and when the computer instructions are executed by a processor, the steps of the laser radar attitude error online measurement and calibration method based on a neural network in the first aspect are implemented.

[0009] The beneficial effects of the present application are:

[0010] The present application realizes error compensation and noise suppression by learning the nonlinear mapping of time series, without the need to construct complex analytical physical models, and the lightweight neural network can directly learn the error correction amount, avoiding explicit decomposition of each error component, has strong generalization ability, is suitable for real-time applications and resource-limited embedded environments, and realizes online measurement and calibration of the attitude parameter error of the laser radar motor system. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is the architecture diagram of the lightweight neural network. DETAILED DESCRIPTION

[0012] The technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0013] It should be understood that when an element is referred to as being "on" another element, it can be directly on the other element or intervening elements can also be present. When an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements can also be present. When an element is referred to as being "fixed" to another element, it can be directly fixed to the other element or intervening elements can also be present.

[0014] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. The terms "comprises", "comprising", "includes", "including" and "have" are inclusive and therefore specify certain disclosed examples, but do not preclude other examples not explicitly described.

[0015] The neural network autonomously designed by the application is a lightweight neural network, which realizes error compensation and noise suppression by learning the nonlinear mapping of time series, does not need to construct a complex analytical physical model, and can directly learn the error correction amount, avoiding explicit decomposition of each error component, is suitable for real-time application and resource-limited embedded environment, and realizes online measurement and calibration of the attitude parameter error of the motor system of the laser radar. The laser radar motor error online measurement and calibration method will be described in detail below in combination with the model structure of the lightweight neural network.

[0016] In a complex environment, the attitude angle data of the laser radar motor w k can be simplified as:

[0017] .

[0018] wherein, A denotes a scale factor and an inter-axis coupling matrix, denotes k the true value of the attitude angle vector of the laser radar motor at time t, b 0 denotes a constant zero offset, b k denotes a random drift, η k denotes a Gaussian white noise, q k denotes a nonlinear error term. In actual application, first, the attitude angle data of the laser radar motor at time t is continuously collected k .Then, the historical attitude angle data with a length of T is input into the lightweight neural network to predict the scale factor w k . T w T . A-1 and nonlinear time-varying error values e k , then based on A -1 , e k , lidar motor k instantaneous pose parameters w k get real-time errors and re-calibrated pose parameter data , the logic of which can be expressed as:

[0019] .

[0020] .

[0021] .

[0022] The focus of the present application is to perform error measurement and construction of a lightweight neural network for calibration. Please refer to Figure 1 , the lightweight neural network adopts a parallel double-branch architecture, which can simultaneously learn global A -1 and e k , to balance global linear correction and local dynamic compensation, improve calibration accuracy and generalization ability. The w k input to the lightweight neural network can be pre-processed to remove noise, and then converted to a standard format for input, which is:

[0023] .

[0024] where R represents the real number field, B represents the batch size, T represents the time series length, D represents the three-dimensional space dimension of the lidar motor pose vector, including pitch angle, roll angle, yaw angle, etc.

[0025] Specifically, the lightweight neural network includes a scale factor learning module, an error estimation module, and an adaptive fusion module. The scale factor learning module is used to learn the systematic scale deviation of the lidar hardware. The error estimation module is used to capture dynamic error patterns in the measurement process. The adaptive fusion module optimally combines the scale deviation and error estimation. The scale factor learning module and the error estimation module are arranged in parallel, and finally output the calibrated through the adaptive fusion module. The and the w kThe same dimension. The following is a lightweight neural network each part of the one-to-one description.

[0026] The scale factor learning module can learn the scale factor in T the mean of the time step sequence statistics, globally correct the scale of the entire sequence, and transform the w k global feature extraction and nonlinear feature transformation processing to obtain A -1 . Specifically, the scale factor learning module includes: a pooling layer, a multilayer perceptron two, a matrix conversion layer, and a product layer. The multilayer perceptron two includes FC layers (fully connected), ReLU function layers, Dropout regularization layers, and FC layers in sequence according to data processing. w k First, enter the pooling layer to perform global average pooling operation in the time dimension, and compress B × T × D dimensional tensor into B × D dimensional pooling feature X global , the calculation formula is:

[0027] .

[0028] wherein, represents the time average value in batch b , feature dimension d . Arrange in batch and feature dimension, that is, obtain the pooling feature X global . X global Enter the first FC layer, so that D dimensional X global is mapped to a 32-dimensional hidden layer, and then sequentially passes through the ReLU function layer and the Dropout regularization layer. The ReLU activation function and the 0.1 Dropout regularization prevent data overfitting. Subsequently, the data after the Dropout regularization layer enters the second FC layer, which maps the 32-dimensional hidden layer feature to a 6-dimensional upper triangular matrix A sym . The matrix conversion layer performs symmetric constraint on the 6-dimensional A sym to meet the inter-axis physical constraint and reduce the learning parameters, and finally reconstructs the 6-dimensional A sym into a 3x3 scale matrix, that is, the required A-1 The specific steps in the matrix conversion process are as follows: let the upper triangular elements be [a11, a12, a13, a22, a23, a33]. Construct a symmetric matrix: A = [[a11, a12, a13], [a12, a22, a23], [a13, a23, a33]]. The scale matrix is A+I, and I represents a 3x3 identity matrix. After obtaining A -1 , the product layer will A -1 perform Hadamard multiplication operation with w k to obtain the scale correction value rec , and the expression is as follows:

[0029] .

[0030] On the other hand, the error estimation module includes a time series learning layer, a multi-head attention mechanism layer, and an error estimation layer. The time series learning layer can extract historical dependent feature information along the time sequence to provide time sequence context information for subsequent error estimation. In this embodiment, the time series learning layer is a unidirectional LSTM (Long Short-Term Memory Network), which sequentially encodes according to time steps. The hidden state of the LSTM contains the local features at this moment and the cumulative information of the previous time steps. Through recursive connection, the LSTM captures the dynamic relationship between each time step in the sequence, including the spatial transformation caused by mechanical transmission, system noise, and time drift effect, etc., to provide time sequence context for error estimation. Finally, the LSTM outputs a time sequence feature with a hidden state dimension of 64 tf . The multi-head attention mechanism layer has an embedding state dimension of 64, uses 4 attention heads, and has a Dropout regularization of 0.1. It applies a multi-head attention mechanism to tf to extract 64-dimensional enhanced features sf , so that the embedded neural network can weight the features of different frames in the time dimension, dynamically strengthen the influence of key time steps, and can integrate global information and enhance features to suppress irrelevant or noise time steps to improve the calibration accuracy. The error estimation layer uses a three-layer perceptron, which can map sf to the real-time error of the predicted measurement value relative to the true value e k for subsequent prediction. The three-layer perceptron for error estimation has the same structure as the two-layer perceptron in the scale factor learning module. It first passes sfDimensionality reduction is performed, and abnormal deviation points in the time series are automatically identified. Then, a ReLU activation function is introduced and a Dropout regularization is used to prevent overfitting. Subsequently, the data is mapped to a 3-dimensional space through a second FC layer e k .

[0031] The adaptive fusion module performs non-linear correction on the output after scale transformation through a residual learning mechanism, compensating for complex error patterns that linear transformation cannot handle. This layer takes advantage of residual connection to add necessary incremental correction to balance scale transformation and error compensation, solving the matching problem between error estimation and scale transformation, and achieving fine adjustment of error estimation with fewer parameters. Specifically, the adaptive fusion module includes a splicing layer one, a multi-layer perception one, and a splicing layer two. The splicing layer one receives rec and e k and performs a splicing operation on the feature dimension to form a fusion feature vector mf The multi-layer perception one is composed of an FC layer, a ReLU function layer, a Dropout regularization layer, and an FC layer in sequence. Among them, the fusion feature vector mf is first compressed through an FC layer, then the ReLU activation function is applied to eliminate negative components, and the Dropout regularization of 0.1 is used to enhance the generalization ability, and finally the FC layer is used to map the compressed scalar feature back to the input dimension. The splicing layer two performs an addition operation on the output of the multi-layer perception one and the scale correction value rec to form the final model output, which is expressed as:

[0032] .

[0033] .

[0034] The inferred based on the lightweight neural network is the laser radar motor attitude angle data after error compensation. The output is subtracted from the laser radar motor system attitude angle data to obtain the real-time error of the laser radar motor system, which is:

[0035] .

[0036] It is worth mentioning that after the embedded neural network is constructed, it needs to be trained before it can be actually used. It can use the conventional training method, and the difference lies in that the total loss function Loss total in the training process is composed of a short-term loss L s calculated based on different time window lengths, a medium-term loss Lm and long-term loss L l The weighted sum is obtained, and its expression is:

[0037] .

[0038] wherein, α , β , γ represent different weight coefficients. In the embodiment, L s , L m , L l The corresponding time window lengths are 16, 32, and 64, respectively. For any kind of loss, the calculation formula is:

[0039] ;

[0040] wherein, N represents the number of time windows, H (·) represents a Huber loss function, represents the estimated value of the attitude angle data in the time window t to t + j , R t，t+j represents the true value of the attitude angle data in the time window t to t + j , j represents the length of the time window.

[0041] In summary, the application proposes a laser radar attitude error online measurement and calibration method based on a lightweight neural network, which realizes real-time measurement and calibration of laser radar scanning data error, constructs an error compensation system with high precision, high robustness and adaptability to variable environments with low computational overhead. Compared with some existing neural network applications in the measurement and navigation fields, such as dilated convolutional neural network (DCNN) and Transformer-based time series modeling, the parameters are too large to calibrate the laser radar motor data in real time in an embedded system, which limits the application in practical engineering. The application can realize online embedded error measurement and parameter self-calibration of the laser radar motor system.

[0042] In another embodiment, a laser radar motor error calibration system based on a lightweight neural network is also proposed, which uses the laser radar attitude error online measurement and calibration method based on a lightweight neural network in the above embodiment. The laser radar motor error calibration system comprises an acquisition module and a correction module. The acquisition module is used to acquire laser radar motork the attitude angle data at the time point w k The correction module is configured to correct the historical attitude angle data of length T w T input into a lightweight neural network to predict a scale factor A -1 and a nonlinear time-varying error value e k based on the scale factor A -1 , e k a laser radar motor k the attitude angle data at the time point w k to obtain real-time error and the attitude angle data after recalibration .

[0043] In another embodiment, a laser radar motor based on an embedded neural network is also proposed, which includes a SoC (System on Chip). The SoC stores a program for implementing the above-mentioned method for online measurement and calibration of laser radar attitude error based on a lightweight neural network. The laser radar motor calls the program to output real-time error and calibrated attitude angle data when working.

[0044] In another embodiment, a computer readable storage medium is also proposed, which stores computer instructions. The computer instructions are executed by a processor to implement the steps of the above-mentioned method for online measurement and calibration of laser radar attitude error based on a lightweight neural network.

[0045] The technical features of the above-mentioned embodiments can be combined in any way. In order to make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described, but as long as the combinations of the technical features do not contradict, they should be considered as the scope of the present disclosure.

[0046] The above-mentioned embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A neural network-based online measurement and calibration method for attitude error of a laser radar, characterized in that, Length of historical pose angle data T w T Input into a light weight neural network to predict scale factor A -1 And non-linear time-varying error value e k Then based on A -1 , e k Lidar motor k Instantaneous pose angle data w k Get real-time error And re-calibrated pose angle data ;​ The light-weight neural network comprises a scale factor learning module, an error estimation module and an adaptive fusion module. The scale factor learning module is configured to learn a scale factor w T perform global feature extraction and nonlinear feature transformation processing to obtain A -1 , and perform multiplication operation on A -1 and w k to obtain a scale correction value rec ; The error estimation module includes: a temporal learning layer, a multi-head attention mechanism layer, and an error estimation layer; the temporal learning layer is used to extract errors sequentially over time. w k Temporal characteristics tf The multi-head attention mechanism layer is used to extract information through a multi-head attention mechanism. tf Enhanced features sf The error estimation layer is used to... sf Nonlinear mapping is e k ; The adaptive fusion module comprises: a splicing layer one, a multi-layer perception one, and a splicing layer two rec and e k The splicing layer one is used for splicing to obtain a fusion feature vector mf The multi-layer perception one is used for learning mf to correct the mapping; and the splicing layer two is used for performing an addition operation on the output of the multi-layer perception one and rec to obtain .

2. The neural network-based online measurement and calibration method for lidar attitude error according to claim 1, characterized in that, w k Preprocessing is performed first, then converted to a standard format and input into a lightweight neural network; the standard format is: ； where R denotes a real field, B denotes a batch size, T denotes a time series length, D denotes a three-dimensional spatial dimension of the lidar motor pose vector.

3. The neural network-based online measurement and calibration of pose error for lidar method according to claim 1, wherein, The scale factor learning module comprises a pooling layer, a multilayer perceptron II, a matrix conversion layer and a product layer. The pooling layer is used for performing global average pooling processing on the feature map to obtain a pooling feature w T The global average pooling processing obtains a pooling feature X global ; Multi-layer perceptron two is used to X global After linear transformation and nonlinear transformation, an upper triangular matrix is obtained A sym ; The matrix conversion layer is used for converting A sym to obtain A -1 ; The product layer is used to multiply A -1 with w k The Hadamard multiplication operation is performed to obtain the scale correction value rec .

4. The neural network-based online measurement and calibration method for lidar attitude error according to claim 3, characterized in that, The multilayer perceptron II comprises, in sequence according to data processing, an FC layer, a ReLU function layer, a Dropout regularization layer and an FC layer.

5. The neural network-based online measurement and calibration of pose error for lidar method according to claim 1, wherein, The temporal learning layer is a unidirectional LSTM that will w k Sequentially encode by time steps and capture the dynamic relationship between w k each time step by recurrent connections.

6. The neural network-based online measurement and calibration of pose error for lidar method according to claim 1, wherein, The light-weight neural network is used after training, the total loss function during training process of which Loss total Short-term loss calculated based on different time window lengths L s Medium-term loss L m And long-term loss L l The result obtained after weighted summation, the expression of which is: ； wherein α , β , γ denote different weight coefficients.

7. The neural network-based online measurement and calibration method for lidar attitude error according to claim 6, characterized in that, L s , L m , L l The calculation formula is: ； wherein, N denotes the number of time windows, H (·) denotes the Huber loss function, denotes a time window t to t + j an estimate of the inner attitude angle data, R t，t+j denotes a time window t to t + j a true value of the inner attitude angle data, j denotes the length of the time window; When computing L s , j takes the value 16; when computing L m , j takes the value 32; when computing L l , j takes the value 64. 8.A neural network based online measurement and calibration system for lidar attitude error, characterized in that, The laser radar attitude error online measurement and calibration method based on the neural network according to any one of claims 1-7 is used. The laser radar attitude error online measurement and calibration method based on the neural network according to any one of claims 1-7 is used. a collection module for collecting laser radar motor k instantaneous attitude angle data w k ; a correction module for correcting historical pose angle data of length T w T into the input calibration neural network to predict a scale factor A -1 and a non-linear time-varying error value e k , then based on A -1 , e k , a lidar motor k pose angle data at time t w k to obtain real-time error and the recalibrated pose angle data .

9. A neural network-based lidar motor, characterized by, The laser radar attitude error online measurement and calibration method based on the neural network according to any one of claims 1-7 is used.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions, and the computer instructions are executed by the processor to implement the steps of the laser radar attitude error online measurement and calibration method based on the neural network according to any one of claims 1-7.

Images