CN-121984515-A - Intelligent driving data compression method, device, vehicle, medium and program product

CN121984515ACN 121984515 ACN121984515 ACN 121984515ACN-121984515-A

Abstract

The embodiment of the application provides an intelligent driving data compression method, device, vehicle, medium and program product, wherein the method comprises the following steps of obtaining a real data set and a coding vector set, wherein the real data set comprises a plurality of real samples of intelligent driving data, and the coding vector set comprises a plurality of coding samples of coding vectors; generating an countermeasure network by utilizing iteration training of a real data set and a coding vector set, wherein the generated countermeasure network comprises a generator and a discriminator, training the discriminator by utilizing a real sample in the real data set, training the generator by utilizing a coding sample in the coding vector set, updating parameters of the generator and the discriminator by utilizing loss in a training process until the total training loss of the generated countermeasure network is smaller than a target loss, and compressing intelligent driving data by utilizing the trained generator. Therefore, the problems that the compression rate is low, key information is easy to lose, the integrity and the usability of data cannot be guaranteed and the like when intelligent driving data are compressed in the related technology are solved.

Inventors

MI CHUNLEI
ZHANG JINGYI
CHEN XIN
LIN DAYANG
YU TENGFEI
GAO YAN

Assignees

北京汽车研究总院有限公司

Dates

Publication Date: 20260505
Application Date: 20251205

Claims (10)

1. An intelligent driving data compression method is characterized by comprising the following steps: acquiring a real data set and a coding vector set, wherein the real data set comprises a plurality of real samples of intelligent driving data, and the coding vector set comprises a plurality of coding samples of coding vectors; Generating an countermeasure network by utilizing the real data set and the code vector set to carry out iterative training, wherein the generating countermeasure network comprises a generator and a discriminator, training the discriminator by utilizing real samples in the real data set, training the generator by utilizing code samples in the code vector set, and updating parameters of the generator and the discriminator by utilizing the loss of a training process until the total training loss of the generating countermeasure network is smaller than a target loss; Intelligent driving data is compressed using the trained generator.
2. The intelligent driving data compression method of claim 1, wherein the generator comprises an encoder and a decoder, the iterative training with the set of real data and the set of encoded vectors generating an countermeasure network, comprising: initializing parameters of the encoder, the decoder and the arbiter; For the current iteration, extracting a real sample from the real data set, encoding the real sample into an encoding vector by the encoder, generating a decoding sample after adding noise to the encoding vector by the decoder, calculating the loss of the discriminator according to the real sample and the decoding sample, and updating the parameters of the discriminator by using the loss of the discriminator; And extracting the coding samples from the coding vector set, generating decoding samples by the decoder after adding noise to the coding samples, calculating the loss of the generator according to the coding samples and the decoding samples, and updating the parameters of the generator by using the loss of the generator.
3. The intelligent driving data compression method according to claim 2, wherein the loss function of the discriminator is: , Wherein, the Representing the loss of the said arbiter, Representing slave decoding distribution Decoding of the mid-samples, Representing a real sample extracted from the real data set, Is a penalty coefficient which is a function of the penalty coefficient, Is a sample obtained by interpolation between a real sample and a decoded sample, A gradient penalty term is represented and, Representing the computation of the arbiter output for a real sample, Representing the computation of the arbiter output for the decoded samples, The gradient of the orientation is represented, Representing a real sample sampled from a real sample distribution, Representing the interpolated samples sampled from the interpolated sample distribution, Representing the desire.
4. The intelligent driving data compression method according to claim 3, wherein, The calculation formula of (2) is as follows: Wherein, the From uniform distribution Data sampled in the memory.
5. The intelligent driving data compression method according to claim 2, wherein the loss function of the generator is: Wherein, the Representing the loss of the generator in question, Representing slave decoding distribution The decoded samples of the mid-samples, Representing the output to the decoded sample generator.
6. The intelligent driving data compression method according to claim 1, wherein the generating a total training loss function of the countermeasure network is: Wherein, the Representing the total training loss of the generated countermeasure network, The super-parameter is represented by a parameter, Representing the loss of the said arbiter, Representing the loss of the generator in question, Representing the loss of an original sample and a decoded sample, wherein the original sample is an input sample before data compression, and the decoded sample is an output sample after compression and decompression; The calculation formula of the loss of the original sample and the decoded sample is as follows: Wherein, the The original sample is represented as such, The corresponding code vector is represented by a code vector, The result of the calculation of the expectation is indicated, Representing the summation symbols, from Added to , Indicating the total number of layers of the arbiter, The normalized weight is represented by a weight of the normalized weight, Representation of the discriminant The feature extraction function of the layer, Representing the L1 norm distance between the two feature vectors.
7. An intelligent driving data compression device, characterized by comprising: The system comprises an acquisition module, a data processing module and a data processing module, wherein the acquisition module is used for acquiring a real data set and a code vector set, the real data set comprises a plurality of real samples of intelligent driving data, and the code vector set comprises a plurality of code samples of code vectors; The training module is used for iteratively training the real data set and the coding vector set to generate an countermeasure network, wherein the countermeasure network comprises a generator and a discriminator, the discriminator is trained by using real samples in the real data set, the generator is trained by using coding samples in the coding vector set, and parameters of the generator and the discriminator are updated by using the loss of a training process until the total training loss of the countermeasure network is smaller than a target loss; and the compression module is used for compressing the intelligent driving data by using the trained generator.
8. A vehicle comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor executing the program to implement the intelligent driving data compression method of any one of claims 1-6.
9. A computer readable storage medium having stored thereon a computer program or instructions, which when executed, implement the intelligent driving data compression method of any of claims 1-6.
10. A computer program product comprising a computer program or instructions which, when executed, implements the intelligent driving data compression method of any one of claims 1-6.

Description

Intelligent driving data compression method, device, vehicle, medium and program product Technical Field The present application relates to the field of intelligent driving technologies, and in particular, to an intelligent driving data compression method, device, vehicle, medium, and program product. Background In an intelligent driving data closed-loop system, a data compression technology is a key technical means for improving data transmission efficiency, reducing storage resource consumption and enhancing data security, and the key aim is to effectively reduce the storage space requirement and transmission bandwidth occupation of data on the premise of ensuring that key information is not lost, so that the overall data processing and transmission efficiency is improved. Because intelligent driving systems need to process data information from a variety of sensors (e.g., cameras, radars, etc.), including video, images, and point clouds, etc., different types of data, both of which involve compression and decompression operations during storage and transmission. However, the compression algorithm in the related art is not specially designed for intelligent driving data, and cannot adapt to compression of different types of data in the intelligent driving data, so that the compression rate is low in the compression process, the occupation of storage space is increased, key information is easily lost, and the integrity and usability of the data cannot be guaranteed. Disclosure of Invention The application provides an intelligent driving data compression method, an intelligent driving data compression device, a vehicle, a medium and a program product, which are used for solving the problems that the compression rate is low, key information is easy to lose, the integrity and the usability of data cannot be ensured and the like when intelligent driving data is compressed in the related technology. The embodiment of the first aspect of the application provides an intelligent driving data compression method which comprises the steps of obtaining a real data set and a coding vector set, wherein the real data set comprises a plurality of real samples of intelligent driving data, the coding vector set comprises a plurality of coding samples of coding vectors, generating an countermeasure network by iterative training of the real data set and the coding vector set, wherein the countermeasure network comprises a generator and a discriminator, training the discriminator by using the real samples in the real data set, training the generator by using the coding samples in the coding vector set, updating parameters of the generator and the discriminator by using losses in a training process until the total training loss of the generated countermeasure network is smaller than a target loss, and compressing the intelligent driving data by using the trained generator. According to the technical means, the real sample training discriminator is used for distinguishing the authenticity of the data, the code sample training generator is used for generating the compressed data, the parameters of the code sample training generator and the code sample training generator are alternately optimized and updated until the total training loss is smaller than the target loss, and the training completed generator is used for carrying out data compression on the intelligent driving data. The method can extract key information of various data in intelligent driving, and different from a general compression algorithm, the method can adjust the compression mode according to different data through continuous optimization of a generator and a discriminator. On the premise of ensuring that key information is not lost, the compression efficiency is improved, the storage cost is saved, the integrity and usability of data are ensured, and the intelligent driving data is efficiently compressed. Optionally, the generator comprises an encoder and a decoder, the counter network is generated by utilizing iteration training of a real data set and a code vector set, the counter network comprises parameters of the encoder, the decoder and the discriminator are initialized, for the current iteration, real samples are extracted from the real data set, the encoder encodes the real samples into code vectors, the decoder adds noise to the code vectors to generate decoding samples, the decoder calculates the loss of the discriminator according to the real samples and the decoding samples, the parameters of the discriminator are updated according to the loss of the real samples, the encoded samples are extracted from the code vector set, the decoder generates decoding samples after adding noise to the encoding samples, the loss of the generator is calculated according to the encoding samples and the decoding samples, and the parameters of the generator are updated according to the loss of the generator. According to the technical means, the embodiment of the app