Convolutional auto-encoder and road feature data compression method based on same

Abstract

The invention provides a convolutional auto-encoder and a road feature data compression method based on the same, and aims to meet the requirement of long-time-domain road feature data compression expression in an automatic driving automobile, and comprehensively extract road plane curvature and vertical section gradient features pre-aiming at a certain mileage through a data compression technology. Meanwhile, the quality of the decompressed data is ensured so as to meet the expression requirements of an automatic driving system on future road forms in the economical driving process, and data acquisition and preprocessing, coordinate system conversion, data segmentation and normalization processing, and building, training and online application of a one-dimensional convolution auto-encoder model are included; the data dimension of the road features can be obviously reduced, the data transmission speed is accelerated, and the communication cost is reduced. Especially for a commercial vehicle, important support is provided for realizing economical cruise control based on the preview road form.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing and compression, and more specifically to a convolutional encoder and a road feature data compression method based on the autoencoder. Background Art

[0002] With the rapid development of self-driving technology, the demand for high-precision maps and road feature data has increased significantly. 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 coordinate system. In particular, through data compression technology, the curvature of the road plane and the slope characteristics of the longitudinal section of a certain mileage can be comprehensively extracted, while ensuring the quality of the decompressed data to meet the needs of the autonomous driving system to express the future road form in the process of achieving economical driving. However, existing data processing methods often face the problem of low storage and transmission efficiency when processing such large-scale and high-dimensional geospatial data. In addition, traditional data compression technology often fails to retain the important geographic and spatial features of the data during the compression process, which is unacceptable for the real-time response and decision-making accuracy of autonomous driving vehicles. Therefore, there is an urgent need for a method that can effectively compress and restore road feature data, which can not only reduce data processing costs, but also ensure data quality and practicality. Summary of the invention

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

[0004] The purpose of the present invention is achieved through the following technical solutions:

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

[0006] Step 1: Collect and pre-process the road elevation and longitude and latitude data required for vehicle autonomous driving through GPS;

[0007] Step 2: Convert the GPS coordinate system collected in step 1 into the UTM coordinate system;

[0008] Step 3: segment the GPS road elevation and the easting coordinates and northing coordinates in the UTM coordinate system, count the maximum and minimum values ​​of each feature, and convert them into data in the vehicle coordinate system;

[0009] Step 4: Normalize the data for different features respectively, and process the data obtained in step 3 into tensor form;

[0010] Step 5: Input the data obtained in step 4 into the convolutional autoencoder that has completed training and testing to process the compression and decompression of the road feature data in real time; the convolutional autoencoder includes an encoding layer and a decoding layer, and the encoding layer and the decoding layer have symmetrical structures. The encoding layer is sequentially composed of five convolutional layers and one fully connected neural network layer, and the decoding layer is sequentially composed of one fully connected layer and five deconvolutional neural networks.

[0011] The encoding layer takes the input vector Encode and decode it to output as a vector

[0012]

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

[0014]

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

[0016] Offline training is to divide the processed data set in step 3 into a training set and a test set; use the training set to train the model, calculate the mean square error to obtain the training target of minimizing the reconstruction error, monitor the learning process and adjust the model parameters;

[0017] The test is to set the convolutional autoencoder to evaluation mode and no longer update any weights. By comparing the test set input data and the reconstructed output data, the reconstruction quality and effect of the model are evaluated until the test set loss function is less than the set target loss function value of 10 -3 .

[0018] As a more optimal technical solution of the present invention: the data preprocessing described in step 1 is achieved by linear interpolation and five-point smoothing.

[0019] As a better technical solution of the present invention: Step 2 specifically determines the UTM zone, converts the longitude and latitude units into radians and calculates the transverse Mercator projection. After the coordinate conversion, the GPS longitude and latitude data are converted into the easting coordinate and the northing coordinate in the UTM coordinate system.

[0020] As a better technical solution of the present invention: Step three specifically segments the GPS road elevation data and the easting coordinates and northing coordinates in the UTM coordinate system, and divides them into training sets and test sets, defines the sliding window size and step size, and then converts the segmented data into the vehicle coordinate system; counts the maximum and minimum values ​​of each feature of the segmented data respectively, in preparation for data normalization.

[0021] As a better technical solution of the present invention: Step four is specifically to normalize the road elevation data and the easting coordinates and northing coordinates in the UTM coordinate system according to the "minimum-maximum normalization" method, and process the normalized training set and test set data into three-dimensional tensors.

[0022] Another object of the present invention is to provide a convolutional autoencoder, comprising

[0023] The encoding layer is used to convolutionally encode the road feature data input into the convolutional autoencoder;

[0024] The decoding layer is used to deconvolute 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 have symmetrical structures. The encoding layer is composed of five convolutional layers and one fully connected neural network layer, and the decoding layer is composed of one fully connected layer and five deconvolutional neural networks. The encoding layer performs a symmetric operation on the input vector. Encode and decode it to output as a vector

[0026]

[0027] f(x) is the encoding function, and g(x) is the decoding function. The output of the encoding function is 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. It is directly used in the vehicle cruise control algorithm to achieve efficient compression of the input data:

[0028]

[0029] In the present invention, the training of the convolutional autoencoder model is carried out completely in an offline environment, with the aim of optimizing the performance of the model and ensuring its high efficiency in practical applications. In the offline training stage, the data set that has been processed is divided into a training set and a test set. The autoencoder is trained using the training set data. The training goal of the autoencoder is to minimize the reconstruction error, which is achieved by calculating the mean square error (MSE), and it is necessary to monitor the learning process and adjust the model parameters, such as adjusting the learning rate and selecting a suitable optimization algorithm, to optimize the model performance. In the test stage, the model is set to evaluation mode and no weights are updated. By comparing the test set input data and the reconstructed 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 value by 10 -3 , which means that the model training is completed and can be used online.

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

[0031] The present invention adopts one-dimensional convolutional autoencoder technology to achieve efficient compression of long-term GPS road feature data. It aims to solve the need for compressed expression of long-term road feature data in autonomous driving vehicles. This method uses effective data compression technology to comprehensively extract road features of a certain mileage, which can significantly reduce the dimension of future road feature data, while ensuring the quality of decompressed data, so as to meet the needs of autonomous driving systems to express future road forms in the process of achieving economical driving goals.

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

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

[0034] Figure 2 It is an overall flow chart of the road feature data compression method based on convolutional autoencoder of the present invention.

[0035] Figure 3 It is the specific network layer design of the convolutional autoencoder model in Example 1 of the present invention.

[0036] Figure 4 It is a trend diagram of the loss function changes of the training set and the test set in Example 1.

[0037] Figure 5 It is a comparison chart of the test set data input and reconstructed output in Example 1. DETAILED DESCRIPTION

[0038] The method of the present invention is further described in detail by the following examples and drawings.

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

[0040] Step 1: Collect and pre-process the road elevation and longitude and latitude data required for vehicle autonomous driving through GPS;

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

[0042] Step 3: segment the GPS road elevation and the easting coordinates and northing coordinates in the UTM coordinate system, count the maximum and minimum values ​​of each feature, and convert them into data in the vehicle coordinate system;

[0043] Step 4: Normalize the data for different features respectively, and process the data obtained in step 3 into tensor form;

[0044] Step 5: Input the data obtained in step 4 into the convolutional autoencoder that has completed training and testing to process the compression and decompression of the road feature data in real time; the convolutional autoencoder includes an encoding layer and a decoding layer, and the encoding layer and the decoding layer have symmetrical structures. The encoding layer is sequentially composed of five convolutional layers and one fully connected neural network layer, and the decoding layer is sequentially composed of one fully connected layer and five deconvolutional neural networks.

[0045] The encoding layer takes the input vector Encode and decode it to output as a vector

[0046]

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

[0048]

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

[0050] Offline training is to divide the processed data set in step 3 into a training set and a test set; use the training set to train the model, calculate the mean square error to obtain the training target of minimizing the reconstruction error, monitor the learning process and adjust the model parameters;

[0051] The test is to set the convolutional autoencoder to evaluation mode and no longer update any weights. By comparing the test set input data and the reconstructed output data, the reconstruction quality and effect of the model are evaluated until the test set loss function is less than the set target loss function value of 10 -3 .

[0052] In some embodiments, the data preprocessing described in step 1 is achieved by linear interpolation and five-point smoothing.

[0053] In some embodiments, step two specifically includes determining the UTM zone, converting the longitude and latitude units into radians and calculating the transverse Mercator projection. After the coordinate conversion, the GPS longitude and latitude data are converted into easting coordinates and northing coordinates in the UTM coordinate system.

[0054] In some embodiments, step three specifically segments the GPS road elevation data and the easting coordinates and northing coordinates in the UTM coordinate system into training sets and test sets, defines the sliding window size and step size, and then converts the segmented data into the vehicle coordinate system; the maximum and minimum values ​​of each feature of the segmented data are counted separately to prepare for data normalization.

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

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

[0057] The encoding layer is used to convolutionally encode the road feature data input into the convolutional autoencoder;

[0058] The decoding layer is used to deconvolute 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 have symmetrical structures. The encoding layer is composed of five convolutional layers and one fully connected neural network layer, and the decoding layer is composed of one fully connected layer and five deconvolutional neural networks. The encoding layer performs a symmetric operation on the input vector. Encode and decode it to output as a vector

[0060]

[0061] f(x) is the encoding function, and g(x) is the decoding function. The output of the encoding function is 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. It is directly used in the vehicle cruise control algorithm to achieve efficient compression of the input data:

[0062]

[0063] Example 1

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

[0065] S1: Data acquisition and preprocessing

[0066] The collected data includes the road elevation data h (unit: m) and latitude and longitude data (unit: degree) required for vehicle autonomous driving collected by GPS equipment. Due to the influence of the environment, 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: Assume that at time t 1 and t 3 With a known height h 1 and h 3 , and at time t 2 The height h 2 is missing, then h 2 It can be obtained by interpolation in the following way:

[0067]

[0068] The data is denoised by five-point smoothing. Take 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] The processed GPS longitude and latitude data are converted into UTM coordinate system data. The specific steps include determining the UTM zone, converting the longitude and latitude units into radians and calculating the transverse Mercator projection. After the coordinate conversion, the GPS longitude and latitude data are converted into the easting coordinate (e) and northing coordinate (n) in the UTM coordinate system.

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

[0073] The GPS road elevation data (h) and the easting coordinates (e) and northing coordinates (n) in the UTM coordinate system are segmented and divided into training sets and test sets. The sliding window size is defined as 100, and the step length of the sliding window is set to 10. Then, the segmented data is converted to the vehicle coordinate system. Taking the elevation data as an example, the relative elevation calculation formula is:

[0074] h i ′=h i -h 0 #(5)

[0075] The calculation of relative easting coordinates and relative northing coordinates is the same. The maximum value h of each feature is counted for the segmented data. imax ′、e imax ′、n imax ′ and minimum value h imin ′、e imin ′、n imin ′, prepare for data normalization.

[0076] S4: Normalize the data for different features and process the data into tensor form

[0077] The three features are normalized according to the "minimum-maximum normalization" method. Taking elevation data as an example, the normalization calculation formula is:

[0078]

[0079] The relative easting coordinates and relative northing coordinates are 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 structures of the encoder and decoder of the one-dimensional convolutional autoencoder in this embodiment are symmetrical. The encoder consists of five convolutional layers and one fully connected neural network layer, and the decoder consists of one fully connected layer and five deconvolutional neural networks. An activation function and a regularization term are added in the middle of each network layer. The specific network layer construction is as follows: Figure 3 As shown. Taking the encoder as an example, the output size of each convolutional layer is calculated as follows:

[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-side edge padding size, and s is the step size. The encoding parameter settings of the convolutional autoencoder are shown in Table 1:

[0084] Table 1

[0085]

[0086]

[0087] The convolutional autoencoder training process is conducted completely offline, and the training goal is to minimize the reconstruction error, that is, the difference between the input data and the output reconstructed by the decoder. This is achieved by calculating the mean square error (MSE) channel by channel and summing them up, so that the model can effectively learn the characteristics of each input channel. The mean square error and specific calculation formula for the three features are:

[0088]

[0089] In addition, in the embodiment, the learning rate is set to 0.001 and the Adam optimization algorithm is selected with 30 iterations to optimize the model performance. In the test phase, the model is set to evaluation mode and no weights are updated. By comparing the input data and the reconstructed 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 of 10 -3 , that is, the model training is completed and can be used online. In Example 1, the mean square error loss function of the training set and the test set changes as follows Figure 4 As shown in the figure, it can be seen that the loss functions of the training set and the test set converge to 10 -3 Below is a comparison chart of the input and reconstructed output of any data selected from the test set ( Figure 5 ), it can be seen that the reconstructed data and the input data have a high degree of overlap, which means that the model training is successful and can be used for online deployment to process the compression and decompression of road feature data in real time.

[0090] The above-mentioned embodiment only expresses one implementation mode of the present invention, and its description is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A road feature data compression method based on convolutional autoencoder, characterized in that: The steps include: Step 1: Collect and pre-process the road elevation and longitude and latitude data required for vehicle autonomous driving through GPS; Step 2: Convert the GPS coordinate system collected in step 1 into the UTM coordinate system; Step 3: segment the GPS road elevation and the easting coordinates and northing coordinates in the UTM coordinate system, count the maximum and minimum values ​​of each feature, and convert them into data in the vehicle coordinate system; Step 4: Normalize the data for different features respectively, and process the data obtained in step 3 into tensor form; Step 5: Input the data obtained in step 4 into the convolutional autoencoder that has completed training and testing to process the compression and decompression of the road feature data in real time; the convolutional autoencoder includes an encoding layer and a decoding layer, and the encoding layer and the decoding layer have symmetrical structures. The encoding layer is sequentially composed of five convolutional layers and one fully connected neural network layer, and the decoding layer is sequentially composed of one fully connected layer and five deconvolutional neural networks. The encoding layer encodes the input vector i and decodes it to output vector i: g(r(i))=i#(1) f(x) is the encoding function, g(x) is the decoding function, and the encoding function outputs a feature vector h, which represents the low-dimensional feature vector extracted from the input data for transmission. It is directly used in the vehicle cruise control algorithm to achieve efficient compression of the input data: f(i)=h#(2).

2. The road feature data compression method based on convolutional autoencoder according to claim 1 is characterized in that: The offline training and testing process of the convolutional autoencoder is as follows: Offline training is to divide the processed data set in step 3 into a training set and a test set; use the training set to train the model, calculate the mean square error to obtain the training target of minimizing the reconstruction error, monitor the learning process and adjust the model parameters; The test is to set the convolutional autoencoder to evaluation mode and no longer update any weights. By comparing the test set input data and the reconstructed output data, the reconstruction quality and effect of the model are evaluated until the test set loss function is less than the set target loss function value of 10 -3 .

3. The road feature data compression method based on convolutional autoencoder according to claim 1 is characterized in that: The data preprocessing described in step 1 is achieved by linear interpolation and five-point smoothing.

4. The road feature data compression method based on convolutional autoencoder according to claim 1 is characterized in that: Step 2 specifically involves determining the UTM zone, converting the longitude and latitude units into radians, and calculating the transverse Mercator projection. After coordinate conversion, the GPS longitude and latitude data are converted into easting coordinates and northing coordinates in the UTM coordinate system.

5. The road feature data compression method based on convolutional autoencoder according to claim 1 is characterized in that: Step three is to segment the GPS road elevation data and the easting coordinates and northing coordinates in the UTM coordinate system, and divide them into training sets and test sets, define the sliding window size and step size, and then convert the segmented data into the vehicle coordinate system; for the segmented data, the maximum and minimum values ​​of each feature are counted respectively to prepare for data normalization.

6. The road feature data compression method based on convolutional autoencoder according to claim 1, characterized in that: Step 4 is to normalize the road elevation data and the easting coordinates and northing coordinates in the UTM coordinate system according to the "minimum-maximum normalization" method, and process the normalized training set and test set data into three-dimensional tensors.

7. A convolutional autoencoder, characterized in that include The encoding layer is used to convolutionally encode the road feature data input into the convolutional autoencoder; The decoding layer is used to deconvolute and reconstruct the road feature data output by the encoding layer, and output the reconstructed road feature data. The encoding layer and decoding layer are symmetrical in structure. The encoding layer is composed of five convolutional layers and one fully connected neural network layer, and the decoding layer is composed of one fully connected layer and five deconvolutional neural networks. The encoding layer encodes the input vector i and decodes it to output vector i: g(f(i))=i#(1) f(x) is the encoding function, and g(x) is the decoding function; the output of the encoding function is 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. It is directly used in the vehicle cruise control algorithm to achieve efficient compression of the input data: f(i)=h#(2).

Images