A convolutional autoencoder and a road feature data compression method based on the same

Abstract

The present application provides a convolutional autoencoder and a road feature data compression method based on the same, aiming to solve the demand for long-time domain road feature data compression expression in an autonomous vehicle, and comprehensively extracts the road plane curvature and longitudinal section slope features within a certain mileage through data compression technology, while ensuring the quality of the decompressed data, to meet the expression demand of the autonomous driving system for the future road form in the process of realizing economic driving, including data acquisition and preprocessing, coordinate system conversion, data segmentation and normalization processing, and the establishment, training and online application of a one-dimensional convolutional autoencoder model, which can significantly reduce the data dimension of road features, accelerate the data transmission speed and reduce the communication cost. In particular for commercial vehicles, it has important support for realizing economic cruise control based on the pre-scan road form.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing and compression, and more particularly to a convolutional encoder and a road feature data compression method based on the same. BACKGROUND

[0002] With the rapid development of autonomous driving technology, there is a significant increase in the demand for high-precision maps and road feature data. These data not only need to be efficiently converted between Global Positioning System (GPS) coordinates and Universal Transverse Mercator (UTM) coordinates, but also need to be accurately processed in the vehicle's coordinate system. In particular, through data compression technology, the road plane curvature and longitudinal profile slope features within a certain mileage can be comprehensively extracted, while ensuring the quality of the decompressed data to meet the needs of the autonomous driving system in expressing the future road form during economic driving. However, existing data processing methods often face the problem of low storage and transmission efficiency when dealing with such large-scale and high-dimensional geospatial data. In addition, traditional data compression techniques often fail to preserve important geographical and spatial features of the data during compression, which is unacceptable for real-time response and decision accuracy of autonomous vehicles. Therefore, there is an urgent need for a method that can effectively compress and restore road feature data, both reducing data processing costs and ensuring data quality and practicality. SUMMARY

[0003] The purpose of the present application is to provide a road feature data compression method based on a convolutional autoencoder, which improves data transmission efficiency and reduces storage requirements through effective data compression technology, while ensuring the quality of the decompressed data to meet the needs of the autonomous driving system for real-time and accurate data processing.

[0004] The purpose of the present application is achieved by the following technical solutions:

[0005] A road feature data compression method based on a convolutional autoencoder, comprising the following steps:

[0006] Step one, collecting and preprocessing the road elevation and latitude and longitude data required for vehicle autonomous driving through GPS;

[0007] Step two, converting the GPS coordinate system collected in step one into the UTM coordinate system;

[0008] Step three, segmenting the GPS road elevation and the eastward and northward coordinates in the UTM coordinate system, respectively, and calculating the maximum and minimum values of each feature, and converting them into data in the vehicle coordinate system;

[0009] Step four, normalizing the data for different features, and processing the data obtained in step three into tensor form;

[0010] Step five, the data obtained in step four is input into the trained and tested convolutional autoencoder for real-time processing of road feature data compression and decompression; the convolutional autoencoder comprises an encoding layer and a decoding layer, the encoding layer and the decoding layer are symmetrical in structure, the encoding layer comprises five convolutional layers and one fully connected layer neural network in sequence, and the decoding layer comprises one fully connected layer and five deconvolutional layers in sequence,

[0011] The encoding layer encodes the input vector and outputs the decoded vector

[0012]

[0013] f(x) is an encoding function, g(x) is a decoding function, the encoding function outputs a feature vector h, which represents a low-dimensional feature vector extracted from the input data for transmission, and is directly used for vehicle cruise automatic control algorithm, realizing efficient compression of input data:

[0014]

[0015] As a more preferred technical solution of the present application: the offline training and testing process of the convolutional autoencoder is as follows:

[0016] Offline training is to divide the data set processed in step three into a training set and a test set; the training set is used to train the model, the minimum reconstruction error training target is obtained by calculating the mean square error, and the model parameters are adjusted during the learning process;

[0017] The test is to set the convolutional autoencoder to evaluation mode without updating any weight, compare the input data and the reconstruction output data of the test set, evaluate the reconstruction quality and effect of the model, and stop until the test set loss function is less than the set target loss function value -3 .

[0018] As a more preferred technical solution of the present application: the data preprocessing in step one is realized by linear interpolation and five-point smoothing.

[0019] As a more preferred technical solution of the present application: step two is specifically to determine the UTM zone, convert the latitude and longitude units to radians and calculate the transverse Mercator projection, and convert the GPS latitude and longitude data to east coordinate and north coordinate in the UTM coordinate system after coordinate conversion.

[0020] As a more optimal technical solution of the present application: step three is specifically to segment the GPS road elevation data and the east coordinate and north coordinate under the UTM coordinate system, and divide them into training set and test set, define the size and step of the sliding window, and then convert the segmented data into the vehicle coordinate system.

[0021] As a more optimal technical solution of the present application: step four is specifically to normalize the road elevation data and the east coordinate and north coordinate under the UTM coordinate system according to the "min-max normalization" method, and process the normalized training set and test set data into three-dimensional tensors.

[0022] The present application also has the purpose of providing a convolutional autoencoder, comprising

[0023] An encoding layer is used to convolve and encode the road feature data input into the convolutional autoencoder;

[0024] A decoding layer is used to deconvolve and reconstruct the road feature data output by the encoding layer, and output the reconstructed road feature data.

[0025] The encoding layer and the decoding layer are symmetrical in structure, the encoding layer is five convolutional layers and one fully connected layer neural network in sequence, the decoding layer is one fully connected layer and five deconvolutional layers neural network in sequence, the encoding layer encodes the input vector And decodes it into a vector

[0026]

[0027] f(x) is the encoding function, g(x) is the decoding function; the encoding function outputs a feature vector h, which represents the potential features extracted from the input data and is also a low-dimensional feature vector used for transmission, which is directly used for vehicle cruise automatic control algorithm to realize efficient compression of input data:

[0028]

[0029] The training of the convolutional autoencoder model in the present application is completely carried out in an offline environment, aiming to optimize the performance of the model and ensure its efficiency in practical application. In the offline training stage, the data set after data processing is divided into a training set and a test set. The autoencoder is trained using the training set data, and the training goal of the autoencoder is to minimize the reconstruction error, which is achieved by calculating the mean square error (MSE), and the learning process needs to be monitored to adjust the model parameters, such as adjusting the learning rate and selecting a suitable optimization algorithm, to optimize the performance of the model. In the test stage, the model is set to evaluation mode and no longer updates any weights. By comparing the test set input data and the reconstruction output data, the reconstruction quality and effect of the model can be intuitively evaluated. Until the test set loss function is less than the set target loss function value10 -3 , it is considered that the model training is completed and can be used online.

[0030] In the automatic driving system, the trained convolutional autoencoder is deployed online to process the compression of road feature data in real time. The high-dimensional (GPS longitude and latitude and elevation) data is compressed into a low-dimensional feature vector using the encoder part, which is used for vehicle cruise automatic control algorithm, effectively reducing the input data dimension of the control algorithm, and also reducing the data transmission and storage requirements. This operation supports the fast and accurate response of the autonomous vehicle under various road conditions, and is a key technology for efficient data processing and instant system response.

[0031] The present application adopts one-dimensional convolutional autoencoder technology to realize efficient compression of long-time domain GPS road feature data. It aims to solve the demand for long-time domain road feature data compression expression in autonomous vehicles. This method can significantly reduce the dimension of future road feature data by effectively compressing the road features within a certain mileage, while ensuring the quality of the decompressed data to meet the demand for expressing future road conditions in the process of achieving economic driving in autonomous driving systems.

[0032] The present application reduces the spatial requirements for storing road feature data required by autonomous vehicles, speeds up data transmission, and reduces communication costs. At the same time, the reconstruction function of one-dimensional convolutional autoencoder in time domain ensures that the compressed data can maintain the original quality after decompression, meeting the accuracy requirements of autonomous driving systems for data, and can be applied to the construction of offline electronic maps and real-time transmission of compressed road information, thereby saving communication resources. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is the overall architecture diagram of the convolutional autoencoder model in the present application.

[0034] Figure 2 is the overall flowchart of the road feature data compression method based on convolutional autoencoder in the present application.

[0035] Figure 3 is the specific network layer design of the convolutional autoencoder model in embodiment 1 of the present application.

[0036] Figure 4 is the loss function trend chart of the training set and test set in embodiment 1.

[0037] Figure 5 is the comparison chart of the test set data input and the reconstruction output in embodiment 1. DETAILED DESCRIPTION

[0038] The method of the present application is further described in detail through the following embodiments and drawings.

[0039] REFERENCE Figure 1 The present application provides a road feature data compression method based on a convolutional autoencoder, comprising the following steps:

[0040] Step 1: Collecting road elevation and latitude and longitude data required for vehicle automatic driving by GPS and preprocessing;

[0041] Step 2: Converting the GPS coordinate system collected in step 1 into a UTM coordinate system;

[0042] Step 3: Segmenting the GPS road elevation and the eastward and northward coordinates in the UTM coordinate system, respectively, and statistically calculating the maximum and minimum values of each feature, and converting them into data in the vehicle coordinate system;

[0043] Step 4: Normalizing the data of different features, and processing the data obtained in step 3 into a tensor form;

[0044] Step 5: Inputting the data obtained in step 4 into the trained and tested convolutional autoencoder for real-time processing of road feature data compression and decompression; the convolutional autoencoder comprises an encoding layer and a decoding layer, the encoding layer and the decoding layer are symmetrical in structure, the encoding layer comprises five convolutional layers and one fully connected layer neural network in sequence, and the decoding layer comprises one fully connected layer and five deconvolutional layers in sequence,

[0045] The encoding layer encodes the input vector and decodes it to output a vector

[0046]

[0047] f(x) is the encoding function, g(x) is the decoding function, the encoding function outputs a feature vector h, which represents a low-dimensional feature vector extracted from the input data for transmission, and is directly used for vehicle cruise automatic control algorithm, realizing efficient compression of input data:

[0048]

[0049] In some embodiments, the offline training and testing process of the convolutional autoencoder is as follows:

[0050] The offline training is to divide the data set processed in step three into a training set and a test set; the training set is used to train the model, the minimum reconstruction error training target is obtained by calculating the mean square error, and the model parameters are adjusted during the learning process;

[0051] The test is to set the convolutional autoencoder to evaluation mode without updating any weight, compare the test set input data and the reconstruction output data, evaluate the reconstruction quality and effect of the model, and stop until the test set loss function is less than the set target loss function value10 -3 .

[0052] In some embodiments, the data preprocessing of step one is achieved by linear interpolation and five-point smoothing.

[0053] In some embodiments, step two is specifically to determine the UTM zone, convert the latitude and longitude units to radians and calculate the transverse Mercator projection, and convert the GPS latitude and longitude data to east coordinate and north coordinate in the UTM coordinate system after coordinate conversion.

[0054] In some embodiments, step three is specifically to segment the GPS road elevation data and the east coordinate and north coordinate in the UTM coordinate system, and divide them into a training set and a test set, define the size and step of the sliding window, and then convert the segmented data into the vehicle coordinate system; the maximum and minimum values of each feature are counted for the segmented data, which is prepared for data normalization.

[0055] In some embodiments, step four is specifically to normalize the road elevation data and the east coordinate and north coordinate in the UTM coordinate system according to the "min-max normalization" method, and process the normalized training set and test set data into three-dimensional tensors.

[0056] The present application also provides a convolutional autoencoder, comprising

[0057] An encoding layer is used to convolve and encode the road feature data input into the convolutional autoencoder;

[0058] A decoding layer is used to deconvolve and reconstruct the road feature data output by the encoding layer, and output the reconstructed road feature data.

[0059] The encoding layer and the decoding layer are symmetrical in structure, the encoding layer comprises five convolutional layers and one fully connected layer neural network in sequence, the decoding layer comprises one fully connected layer and five deconvolutional layers in sequence, and the encoding layer is used to calculate the input vector Encoding and decoding output as a vector

[0060]

[0061] f(x) is the encoding function, g(x) is the decoding function; the encoding function outputs a feature vector h, which represents the potential features extracted from the input data, and is also a low-dimensional feature vector used for transmission, which is directly used for vehicle cruise automatic control algorithm to realize efficient compression of input data:

[0062]

[0063] Example 1

[0064] Referring to Figure 2 The data processing and the specific process of building and offline training of the one-dimensional convolutional autoencoder in the embodiment are shown in the figure, and the following are the specific implementation steps.

[0065] S1: Data acquisition and preprocessing

[0066] The collected data includes road elevation data h (unit: m) and latitude and longitude data (unit: degree) required for vehicle automatic driving collected by GPS equipment. Due to environmental factors, there are missing values, noise and abnormal data points in the data. The data needs to be preprocessed, and the missing values are processed by linear interpolation. Taking the elevation data as an example: assuming that the elevations h1 and h3 at times t1 and t3 are known, and the elevation h2 at time t2 is missing, then h2 can be obtained by interpolation as follows:

[0067]

[0068] And the data is denoised by five-point smoothing. Taking the elevation data as an example: assuming that the data at position j is smoothed, the five-point smoothing formula is:

[0069]

[0070] S2: Convert GPS coordinate system to UTM coordinate system

[0071] Convert the processed GPS latitude and longitude data into UTM coordinate system data. The specific steps include determining the UTM zone, converting the latitude and longitude units to radians and calculating the horizontal Mercator projection. After coordinate conversion, the GPS latitude and longitude data is converted into east coordinate (e) and north coordinate (n) in the UTM coordinate system.

[0072] S3: Set sliding window to segment data and convert to data in vehicle coordinate system

[0073] GPS road elevation data (h) and UTM coordinate system eastward coordinate (e), northward coordinate (n) are segmented and divided into training set and test set. The size of the sliding window is defined as 100, the step size of the sliding window is set as 10, then the segmented data is converted into the vehicle coordinate system, taking the elevation data as an example, the relative elevation calculation formula is:

[0074] h i ′=h i -h0#(5)

[0075] The relative eastward coordinate and the relative northward coordinate are also calculated. The maximum value h imax ′、e imax ′、n imax ′ and the minimum value h imin ′、e imin ′、n imin ′ of each feature of the segmented data are counted respectively, which is prepared for data normalization.

[0076] S4: Data normalization is performed on different features respectively, and the data is processed into tensor form

[0077] According to the "minimum-maximum normalization" method, three features are normalized respectively, taking the elevation data as an example, the normalization calculation formula is:

[0078]

[0079] The relative eastward coordinate and the relative northward coordinate are also normalized by the same method. And the normalized training set and test set data are processed into a three-dimensional tensor of m*3*100.

[0080] S5: Build a convolutional autoencoder model and perform offline model training, testing and parameter adjustment

[0081] The structure of the encoder part and the decoder part of the one-dimensional convolutional autoencoder in this embodiment is symmetrical, the encoder is composed of five convolutional layers and one fully connected layer neural network, the decoder is composed of one fully connected layer and five deconvolutional layer neural networks, an activation function and a regularization term are added in each network layer, and the specific network layer construction is shown in Figure 3 . Taking the encoder as an example, the output size of each convolutional layer is calculated:

[0082]

[0083] In the formula, l' is the output size, l is the input size, k is the convolution kernel size, p is the single-sided edge padding size, and s is the step size. The parameter settings of the encoding part of the convolutional autoencoder are shown in Table 1:

[0084] Table 1

[0085]

[0086]

[0087] The convolutional autoencoder training process is completely carried out in an offline environment, and the training target is to minimize the reconstruction error, that is, the difference between the input data and the output reconstructed by the decoder. The mean square error (MSE) is calculated channel by channel and summed up, so that the model can effectively learn the characteristics of each input channel. The mean square error and specific calculation formula for three features are as follows:

[0088]

[0089] In addition, in the embodiment, the learning rate is set to 0.001 and the Adam optimization algorithm is selected, and the iteration number is 30, so as to optimize the performance of the model. In the test stage, the model is set to evaluation mode and no longer updates any weight. By comparing the input data and the reconstruction output, the reconstruction quality and effect of the model can be intuitively evaluated. Until the test set loss function is less than the set target loss function value 10 -3 , it is considered that the model training is completed and can be used online. In embodiment 1, the mean square error loss function changes of the training set and the test set are as shown in Figure 4 The figure shows that the loss functions of the training set and the test set converge to 10 -3 below. From the comparison chart of the input and the reconstruction output of any data selected from the test set Figure 5 , it can be seen that the reconstructed data and the input data have high coincidence degree, that is, the model training is successful and can be used for online deployment to realize real-time compression and decompression of road feature data.

[0090] The above-described embodiments only express one embodiment of the present application, which is described in detail and specifically, but should not be understood as limiting the scope of the present application. 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 all within the scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1.A method for compressing road feature data based on a convolutional autoencoder, characterized in that, Comprising the following steps: Step one, preprocessing by GPS collecting the road elevation and latitude and longitude data required for automatic driving of the vehicle; Step two, converting the preprocessed latitude and longitude data into UTM coordinate system; specifically, determining the UTM zone, converting the latitude and longitude units to radians and calculating the Mercator projection of the horizontal axis, and converting the GPS latitude and longitude data into eastward and northward coordinates in the UTM coordinate system after coordinate conversion; Step three, segmenting the GPS road elevation and eastward and northward coordinates in the UTM coordinate system, and converting them into data in the vehicle coordinate system; Step four, respectively, the maximum and minimum values of each feature are counted, the data of different features are normalized, and the obtained data are processed into a three-dimensional tensor form; Step five, inputting the data obtained in step four into the trained and tested convolutional autoencoder for real-time processing of road feature data compression and decompression; the convolutional autoencoder comprises an encoding layer and a decoding layer, the encoding layer and the decoding layer are symmetrical in structure, the encoding layer comprises five convolutional layers and one fully connected layer neural network in sequence, and the decoding layer comprises one fully connected layer and five deconvolutional layers in sequence; Encoding layer for input vectors i Encoding and decoding output as vectors i ＃ : g( f ( i )）= i ＃ ; f i ) encoding function, f i h g () is a decoding function, the output of the encoding function is a feature vector h , which represents the low-dimensional feature vector extracted from the input data for transmission, which is directly used in the vehicle cruise automatic control algorithm to achieve efficient compression of the input data.​​​​ 2.The road feature data compression method based on a convolutional auto-encoder according to claim 1, characterized in that, The offline training and testing process of the convolutional autoencoder is as follows: Offline training is to divide the data set processed in step three into a training set and a test set; the training set is used to train the model, the minimum reconstruction error training target is obtained by calculating the mean square error, and the model parameters are adjusted during the learning process; The test is to set the convolutional autoencoder to evaluation mode without updating any weight, compare the input data and the reconstructed output data of the test set, and evaluate the reconstruction quality and effect of the model. 3.The road feature data compression method based on a convolutional auto-encoder according to claim 1, characterized in that, The data preprocessing of step one is realized by linear interpolation and five-point smoothing. 4.The road feature data compression method based on a convolutional auto-encoder according to claim 1, wherein, Step three is specifically to define the size and step of the sliding window, segment the GPS road elevation data and eastward and northward coordinates in the UTM coordinate system, and divide them into a training set and a test set, and then convert the segmented data into the vehicle coordinate system; the maximum and minimum values of each feature are counted for the segmented data, and the data normalization is prepared.

Images