A high-dimensional lossy source coding method and device for automatic driving

Abstract

The application provides a high-dimensional lossy source decoding method and device for automatic driving, in which n-dimensional source data from a vehicle-mounted sensor is input into a lossy source encoder network to obtain a k-dimensional quaternary sequence; based on a check matrix of a pre-stored original pattern low-density parity check code, algebraic reconstruction is performed on the k-dimensional quaternary sequence on an integer ring to generate an n-dimensional code word satisfying a check constraint; the n-dimensional code word is centrally encoded to map discrete quaternary symbols therein into a continuous numerical sequence centered on zero; and the continuous numerical sequence is input into a pre-trained deep decoder network to obtain a reconstructed Gaussian floating point sequence for environmental perception of automatic driving. Therefore, the scheme provided in the embodiments of the application effectively improves the decoding efficiency and data reconstruction quality in an automatic driving system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information processing, and relates to communication physical layer decoding technology, and in particular to a high-dimensional lossy source decoding method and device for automatic driving. BACKGROUND

[0002] High-dimensional lossy source coding technology shows important value in multiple fields. In the automatic driving system, efficient compression transmission and accurate reconstruction of high-dimensional point cloud data generated by sensors such as LiDAR are the core links to realize safe and reliable environmental perception. This technology aims to utilize the inherent sparsity and correlation of high-dimensional data to significantly reduce the data load of the vehicle network and the storage pressure of the computing unit on the premise of ensuring the fidelity of key features.

[0003] Currently, D. Song et al. proposed a lossy source coding system based on the cascade belief propagation-inverse belief propagation algorithm in “Gaussian Source Coding Based on P-LDPC Code”, which optimizes the efficient compression and reconstruction of Gaussian sources. The lossy source decoding network adopts a fully connected layer structure, which can reconstruct the compressed binary sequence into a Gaussian floating-point source. Compared with traditional methods, this decoding scheme has the advantages of simple implementation, low iteration complexity, and easy parallel processing, providing a new technical path for efficient information representation and reconstruction.

[0004] However, in real-time applications such as automatic driving that are critical to safety, the compressed data stream received from the vehicle network must be recovered to the original sensor readings quickly and accurately, while the current lossy source decoding method mainly reconstructs binary information, which has bottlenecks in information density and decoding efficiency.

[0005] Therefore, how to solve the efficient and accurate reconstruction of compressed sensor data stream in the automatic driving system is a problem that needs to be solved by those skilled in the art. SUMMARY

[0006] The embodiments of the present application provide a high-dimensional lossy source decoding method and device for automatic driving, which can effectively improve the decoding efficiency and data reconstruction quality in the automatic driving system. Compared with the original binary representation, the high-dimensional data compression provided by the present application can represent the multi-dimensional information of automatic driving, realize efficient multi-target information representation, and provide a new technical path for the identification scheme of automatic driving.

[0007] The first aspect of the present application provides a high-dimensional lossy source decoding method for automatic driving, comprising:

[0008] Inputting n-dimensional source data from vehicle-mounted sensors into a lossy source encoder network to obtain a k-dimensional quaternary sequence;

[0009] Based on the pre-stored original graph low-density parity-check code check matrix, in the integer ring Algebraic reconstruction is performed on the k-dimensional quaternary sequence to generate an n-dimensional code word satisfying the check constraint;

[0010] Centralized encoding is performed on the n-dimensional code word to map the discrete quaternary symbols therein to a continuous numerical sequence centered on zero;

[0011] The continuous numerical sequence is input into a pre-trained deep decoder network to obtain a reconstructed Gaussian floating-point sequence for environmental perception by an autonomous vehicle.

[0012] Optionally, based on the pre-stored original graph low-density parity-check code check matrix, in the integer ring Algebraic reconstruction is performed on the k-dimensional quaternary sequence to generate an n-dimensional code word satisfying the check constraint includes:

[0013] Determining an information bit index according to configuration information of the original graph low-density parity-check code;

[0014] Filling the received quaternary sequence into a position specified by the information bit index in a code word vector of length n;

[0015] Based on the row echelon form of the check matrix, all check bits of the code word vector are calculated and filled by a back substitution method containing modulo 4 operation to generate an n-dimensional code word satisfying the check constraint.

[0016] Optionally, the process of calculating the check bit value is implemented by the following formula:

[0017]

[0018] wherein, represents the value of the current check bit to be solved, represents an element in the i-th row and j-th column of the row echelon form check matrix, represents other determined bits in the code word vector except the current check bit to be solved; represents a modulo 4 operation.

[0019] Optionally, before the centralized encoding is performed on the n-dimensional code word, the method further includes:

[0020] Verifying whether the n-dimensional code word satisfies the following check equation:

[0021]

[0022] denotes converting an n-dimensional code word from a row vector to a column vector denotes a row staircase check matrix; denotes a modulo 4 operation.

[0023] Optionally, the deep decoder network comprises an input layer, a hidden layer, and an output layer, wherein:

[0024] The input layer maps the sequence of continuous numerical values to a 4n-dimensional hidden space through a linear transformation;

[0025] The hidden layer employs a GELU activation function;

[0026] The output layer maps the 4n-dimensional representation of the hidden layer back to the n-dimensional reconstructed source.

[0027] Optionally, the weights of the input layer, the hidden layer, and the output layer of the deep decoder network are initialized with Xavier uniform distribution.

[0028] Optionally, the centering encoding maps the quaternary symbols {0, 1, 2, 3} to {-1.5, -0.5, 0.5, 1.5}, respectively.

[0029] The second aspect of the present application provides a high-dimensional lossy source decoding device for autonomous driving, which is arranged in a vehicle-mounted communication system and used for reconstructing point cloud data collected by a sensor, and comprises:

[0030] A quantization unit is configured to quantize n-dimensional source data from a vehicle-mounted sensor into a k-dimensional quaternary sequence through a lossy source encoder network;

[0031] An algebraic reconstruction unit is configured to perform algebraic reconstruction on the k-dimensional quaternary sequence on an integer ring based on a check matrix of a pre-stored original graph low-density parity-check code, to generate an n-dimensional code word satisfying a check constraint;

[0032] A centering unit is configured to perform centering encoding on the n-dimensional code word, to map discrete quaternary symbols in the n-dimensional code word into a continuous numerical sequence centered on zero;

[0033] A decoding unit is configured to input the continuous numerical sequence into a pre-trained deep decoder network, to obtain a reconstructed Gaussian floating-point sequence for environmental perception by autonomous driving.

[0034] The third aspect of the present application provides a high-dimensional lossy source decoding device for autonomous driving, comprising:

[0035] One or more processors;

[0036] A memory having one or more programs stored thereon;

[0037] The one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method for high-dimensional lossy source coding for autonomous driving according to any one of the above.

[0038] The fourth aspect of the present application provides a computer storage medium for storing a program, which, when executed, implements the method for high-dimensional lossy source coding for autonomous driving according to any one of the above.

[0039] Compared with the prior art, the beneficial effects of the present application include:

[0040] Improving information density and coding efficiency: The present application uses quaternary coding, each symbol can carry 2 bits of information, and the information density is twice that of binary coding. When processing the same amount of data, the number of symbols required for processing is halved, significantly reducing the complexity of the coding algorithm, improving the processing speed and efficiency, and meeting the stringent requirements of real-time applications such as autonomous driving.

[0041] Improving the quality of reconstructed signals: Quaternary coding provides more refined quantization levels, which can effectively reduce quantization noise compared to binary schemes, improving the accuracy of reconstructed data and providing a more reliable data foundation for subsequent environmental perception algorithms.

[0042] Combining the advantages of algebra and deep learning: The present application innovatively combines the algebraic reconstruction characteristics of protograph low-density parity-check (P-LDPC) codes with the powerful nonlinear mapping capabilities of deep learning. First, the legality of the code word is guaranteed by algebraic methods, and then the complex mapping from the code word to the source is learned using a neural network, fully utilizing the advantages of both, achieving better reconstruction performance under the same network capacity.

[0043] Strong engineering practicability: The method and device for high-dimensional lossy source coding for autonomous driving of the present application can be directly integrated into existing vehicle-mounted communication units or edge computing devices, without the need to modify the hardware infrastructure, and has good implementability and application prospects. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments described in the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0045] Figure 1 A flowchart of a method for high-dimensional lossy source coding for autonomous driving according to an embodiment of the present application is provided.

[0046] Figure 2 A processing flow schematic diagram of the lossy source coding provided by the embodiment of the present application is provided.

[0047] Figure 3 A structure schematic diagram of the high-dimensional lossy source coding device for automatic driving provided by the embodiment of the present application is provided.

[0048] Figure 4 A structure schematic diagram of the high-dimensional lossy source coding device for automatic driving provided by the embodiment of the present application is provided. DETAILED DESCRIPTION

[0049] The embodiment of the present application provides a high-dimensional lossy source coding method and device for automatic driving, which accurately reconstructs a compressed quaternary bit point into an original floating point number, and provides a data reconstruction scheme with higher coding efficiency, and reconstructs a compressed quaternary data sequence received from a vehicle-mounted network into a Gaussian distributed floating point number sequence, so that the coding efficiency and data reconstruction quality of an automatic driving system are effectively improved.

[0050] Compared with an existing binary coding scheme, the present technology provides a high-dimensional lossy source coding method based on quaternary. Since the information amount carried by each symbol of quaternary is twice that of binary, the number of symbols required to be processed by the decoder when processing the same information content is halved, which significantly improves the convergence speed and processing efficiency of the coding algorithm. Especially in the scene of LiDAR intensity value which is sensitive to quantization noise, the reconstruction based on quaternary can provide more fine posterior information, and increase the peak signal-to-noise ratio of the reconstructed signal by 2-4dB, which is crucial for subsequent target detection and recognition algorithms.

[0051] Referring to Figure 1 , the figure is a flow schematic diagram of the high-dimensional lossy source coding method for automatic driving provided by the embodiment of the present application. The high-dimensional lossy source coding method for automatic driving provided by the embodiment of the present application can be realized, for example, through the following steps S101-S104.

[0052] The embodiment of the present application is described in combination with Figure 2 Figure 2 ​The processing flow diagram of the lossy source coding provided by the embodiment of the present application shows the module logic and data flow from compressed data to reconstructed source: first, the input is a compressed quaternary sequence; then, through algebraic reconstruction of the code word, a complete code word satisfying the check constraint is generated based on the check matrix of the P-LDPC code; then, through the central encoding, the quaternary code word is mapped to the zero-centered continuous numerical value (for example, the quaternary symbol {0, 1, 2, 3} is mapped to {-1.5, -0.5, 0.5, 1.5}); the dashed box is the "deep decoder", which includes a multi-layer network structure of input layer→hidden layer→output layer, and performs nonlinear transformation on the centralized numerical value; finally, the reconstructed Gaussian floating point sequence (restored to the original high-dimensional floating point number source of LiDAR, used for automatic driving environment perception) is output.

[0053] S101: input the n-dimensional source data from the vehicle-mounted sensor into the lossy source encoder network to obtain a k-dimensional quaternary sequence.

[0054] In the embodiment of the present application, the n-dimensional floating point data from the vehicle-mounted LiDAR sensor, which approximately obeys the standard normal distribution, is input into the lossy source encoder network to be quantized into a k-dimensional quaternary compressed sequence u, wherein the dimension satisfies n≥k.

[0055] Specifically, the lossy source encoder network adopts a deep fully connected structure, specifically including an input layer, a feature extraction layer and a quantization output layer, for compressing and encoding the n-dimensional floating point source data into a k-dimensional quaternary sequence. The input layer receives the n-dimensional floating point data from the vehicle-mounted LiDAR sensor, which approximately obeys the standard normal distribution; the feature extraction layer includes 3 fully connected layers, which are nonlinearly transformed by GELU activation function between layers, the first fully connected layer maps the n-dimensional input to a 2n-dimensional feature space, the second fully connected layer maps the 2n-dimensional feature to an n-dimensional feature space; the third fully connected layer maps the n-dimensional feature to a k-dimensional feature space; the quantization output layer performs quaternary quantization on the continuous feature, uses a learnable threshold to quantize the continuous value into a quaternary compressed sequence u, and uses a uniform quantization strategy for the quantization function, and the quantization step is adaptively adjusted according to the statistical characteristics of the source.

[0056] S102: based on the check matrix of the pre-stored original graph low-density parity check code, algebraically reconstruct the k-dimensional quaternary sequence in the integer ring to generate an n-dimensional code word satisfying the check constraint.

[0057] In the embodiment of the present application, the information bit index is determined according to the configuration information of the original graph low-density parity check code; the received quaternary sequence is filled into the position specified by the information bit index in the code word vector with a length of n; based on the row echelon form of the check matrix, all the check bits of the code word vector are calculated and filled by the reverse substitution method containing modulo 4 operation, to generate an n-dimensional code word satisfying the check constraint.

[0058] Specifically, an empty codeword vector with length n is created, and the compressed sequence is filled into the vector at positions determined by the information bit indexes cfg.kept_indices of the P-LDPC code, at which time the information bits are determined, and the check bits are to be solved. Based on the given P-LDPC code, a row echelon form check matrix is realized, and all check bits are calculated and filled by using a back substitution method.

[0059] Each check bit is processed in turn in the reverse order of the check bit indexes; for the current check bit to be calculated, only one unknown quantity is contained in the check equation corresponding to the check bit, and other items are determined; the value of the current check bit is solved by calculating the dot product of the check equation and the partially filled codeword vector, and performing a modulo 4 operation on the result, and the calculation formula is represented as:

[0060]

[0061] wherein, represents the value of the check bit to be solved, denotes an element in the i-th row and the j-th column of the row echelon form check matrix, is a determined bit in the codeword vector other than the current check bit to be solved; denotes a modulo 4 operation.

[0062] The calculated check bit value is filled into the corresponding position of the codeword vector, and the above calculation process is repeated until all check bits are calculated.

[0063] In an implementation manner of an embodiment of the present application, after the algebraic reconstruction is completed, the generated codeword is represented as Meanwhile, the legality is verified based on the check matrix of the P-LDPC code, to ensure that it satisfies the check equation, which is represented as:

[0064]

[0065] wherein, represents that the n-dimensional codeword is converted from a row vector to a column vector represents the row echelon form check matrix; denotes a modulo 4 operation. After the verification is passed, step S103 is entered.

[0066] S103: The n-dimensional codeword is centrally coded, and the discrete quaternary symbol therein is mapped to a continuous numerical sequence centered on zero.

[0067] In an embodiment of the present application, the codeword is centrally coded, that is, ​The four symbols {0, 1, 2, 3} are mapped to {-1.5, -0.5, 0.5, 1.5} respectively, and the centralization encoding is used to convert the discrete quaternary sequence into a zero-centered continuous numerical sequence, so that the decoder network can learn the mapping relationship from the code word to the source more effectively.

[0068] S104: input the continuous numerical sequence into the pre-trained deep decoder network to obtain a reconstructed Gaussian floating point sequence for environment perception of autonomous driving.

[0069] In the embodiment of the application, the deep decoder network comprises an input layer, a hidden layer and an output layer, wherein: the input layer maps the continuous numerical sequence to a 4n-dimensional hidden space through linear transformation; the hidden layer adopts a GELU activation function; and the output layer maps the 4n-dimensional representation of the hidden layer back to the n-dimensional reconstructed source. The weights of the input layer, the hidden layer and the output layer of the deep decoder network are initialized using Xavier uniform distribution.

[0070] Specifically, referring to Figure 2 The reconstructed centralization code word is input into the deep decoder network . The network structure mainly comprises an input layer, a hidden layer and an output layer. First, the input layer maps the n-dimensional centralization reconstructed code word to a 4n-dimensional hidden space through linear transformation; second, the hidden layer adopts a GELU activation function to provide nonlinear transformation capability; and then, the output layer maps the 4n-dimensional representation of the hidden layer back to the n-dimensional reconstructed source . The weights of each layer are initialized using Xavier uniform distribution.

[0071] The application provides a high-dimensional lossy source decoding method and device for autonomous driving. In the method, n-dimensional source data from a vehicle-mounted sensor is input into a lossy source encoder network to obtain a k-dimensional quaternary sequence; based on a pre-stored check matrix of a protograph low-density parity-check code, the k-dimensional quaternary sequence is algebraically reconstructed on an integer ring to generate an n-dimensional code word satisfying a check constraint; the n-dimensional code word is subjected to centralization encoding to map discrete quaternary symbols therein to a zero-centered continuous numerical sequence; and the continuous numerical sequence is input into a pre-trained deep decoder network to obtain a reconstructed Gaussian floating point sequence for environment perception of autonomous driving. As can be seen, the scheme provided in the embodiment of the application reconstructs the compressed quaternary sequence into a floating point sequence conforming to a standard normal distribution, learns the source features while fully utilizing the algebraic properties of the code word, and can achieve better reconstruction performance under the same network capacity, thereby effectively improving the decoding efficiency and data reconstruction quality in the autonomous driving system.

[0072] Based on the method provided in the above embodiments, the embodiment of the present application further provides a high-dimensional lossy source coding device for automatic driving.

[0073] Referring to Figure 3 The figure is a structural schematic diagram of a high-dimensional lossy source coding device for automatic driving provided by the embodiment of the present application.

[0074] The high-dimensional lossy source coding device for automatic driving 300 provided by the embodiment of the present application comprises a quantization unit 301, an algebraic reconstruction unit 302, a centralization unit 303 and a decoding unit 304.

[0075] The quantization unit 301 is configured to quantize n-dimensional source data from a vehicle-mounted sensor into a k-dimensional quaternary sequence in a lossy source encoder network;

[0076] The algebraic reconstruction unit 302 is configured to perform algebraic reconstruction on the k-dimensional quaternary sequence on an integer ring based on a check matrix of a pre-stored original graph low-density parity-check code, to generate an n-dimensional code word satisfying a check constraint;

[0077] The centralization unit 303 is configured to perform centralization coding on the n-dimensional code word, and map discrete quaternary symbols therein into a continuous numerical sequence centered on zero;

[0078] The decoding unit 304 is configured to input the continuous numerical sequence into a pre-trained deep decoder network, to obtain a reconstructed Gaussian floating-point sequence, for environment perception of automatic driving.

[0079] In a possible implementation, the algebraic reconstruction unit 302 is specifically configured to:

[0080] determine an information bit index according to configuration information of the original graph low-density parity-check code;

[0081] fill the received quaternary sequence into a position specified by the information bit index in a code word vector with a length of n;

[0082] based on a row echelon form of the check matrix, calculate and fill all check bits of the code word vector through a back substitution method containing a modulo 4 operation, to generate an n-dimensional code word satisfying the check constraint.

[0083] In a possible implementation, the algebraic reconstruction unit 302 is specifically configured to:

[0084]

[0085] wherein, represents a value of a check bit to be solved at present, denotes an element of the i-th row and the j-th column of the row echelon form check matrix, is a certain position other than the current to-be-solved check bit in the code word vector. is a certain position other than the current to-be-solved check bit in the code word vector. represents a modulo 4 operation.

[0086] In a possible implementation, the high-dimensional lossy source coding apparatus for automatic driving 300 further includes a check unit, configured to:

[0087] verify whether the n-dimensional code word satisfies the following check equation:

[0088]

[0089] wherein, represents converting the n-dimensional code word from a row vector to a column vector represents a row echelon check matrix. represents a modulo 4 operation.

[0090] In a possible implementation, the deep decoder network includes an input layer, a hidden layer, and an output layer, wherein:

[0091] the input layer maps the sequence of continuous numerical values to a 4n-dimensional hidden space through linear transformation;

[0092] the hidden layer adopts a GELU activation function.

[0093] the output layer maps the 4n-dimensional representation of the hidden layer back to the n-dimensional reconstructed source.

[0094] In a possible implementation, the weights of the input layer, the hidden layer, and the output layer of the deep decoder network are all initialized using Xavier uniform distribution.

[0095] In a possible implementation, the centralized encoding maps the quaternary symbols {0, 1, 2, 3} to {-1.5, -0.5, 0.5, 1.5}, respectively.

[0096] Since the high-dimensional lossy source coding apparatus for automatic driving 300 is a device corresponding to the high-dimensional lossy source coding method for automatic driving provided in the above method embodiments, the specific implementation of each unit of the high-dimensional lossy source coding apparatus for automatic driving 300 is the same as the same concept in the above method embodiments, and therefore, for the specific implementation of each unit of the high-dimensional lossy source coding apparatus for automatic driving 300, reference can be made to the description of the high-dimensional lossy source coding method for automatic driving in the above method embodiments, which will not be repeated here.

[0097] The embodiment of the application further provides a high-dimensional lossy source coding device for automatic driving, the device comprising: a processor and a memory.

[0098] The memory is configured to store instructions.

[0099] The processor is configured to execute the instructions in the memory, and execute the high-dimensional lossy source coding method for automatic driving performed by an analysis device mentioned in the above embodiments.

[0100] It should be noted that the hardware structure of the high-dimensional lossy source coding device for automatic driving provided in the embodiments of the present application can be as shown in Figure 4 Figure 4 The structure of a device provided in the embodiments of the present application is shown in the figure.

[0101] Referring to Figure 4 , the device 400 includes a processor 410, a communication interface 420 and a memory 430. The number of processors 410 in the device 400 can be one or more, Figure 4 In the embodiment of the present application, the processor 410, the communication interface 420 and the memory 430 can be connected through a bus system or other means, wherein, Figure 4 In the embodiment of the present application, the processor 410, the communication interface 420 and the memory 430 can be connected through a bus system or other means, wherein,

[0102] The processor 410 can be a central processing unit (CPU), a network processor (NP) or a combination of CPU and NP. The processor 410 can further include a hardware chip. The hardware chip can be an application-specific integrated circuit (ASIC), a programmable logic device (PLD) or a combination thereof. The PLD can be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL) or any combination thereof.

[0103] ​The memory 430 can include a volatile memory (e.g., random-access memory (RAM)), and can also include a non-volatile memory (e.g., flash memory, a hard disk drive (HDD), or a solid-state drive (SSD)). The memory 430 can also include a combination of the above-mentioned types of memories.

[0104] Optionally, the memory 430 stores an operating system and a program, an executable module, or a data structure, or a subset thereof, or an extended set thereof, wherein the program can include various operation instructions for implementing various operations. The operating system can include various system programs for implementing various basic services and processing hardware-based tasks. The processor 410 can read the program in the memory 430 to implement the high-dimensional lossy source coding method for autonomous driving provided by the embodiments of the present application.

[0105] The bus system 440 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus system 440 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 Only one thick line is used in the figure to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0106] The embodiments of the present application also provide a computer readable storage medium, including instructions, which, when executed on a computer, cause the computer to perform the high-dimensional lossy source coding method for autonomous driving mentioned in the above embodiments.

[0107] The embodiments of the present application also provide a computer program product including instructions, which, when executed on a computer, cause the computer to perform the high-dimensional lossy source coding method for autonomous driving mentioned in the above embodiments.

[0108] Although the present application is specifically shown and described in connection with the preferred embodiments, those skilled in the art should understand that various changes in form and details can be made to the present application without departing from the spirit and scope of the present application as defined in the appended claims, and all such changes are intended to be within the protection scope of the present application.

Claims

1. A high-dimensional lossy source coding method for autonomous driving, characterized in that, The method is executed by a vehicle-mounted communication system, and is used for reconstructing point cloud data collected by a sensor, and comprises the following steps of: inputting n-dimensional source data from a vehicle-mounted sensor into a lossy source encoder network to obtain a k-dimensional quaternary sequence; Based on the pre-stored original mold map low density parity check code check matrix, in the integer ring The k-dimensional quaternary sequence is algebraically reconstructed, and an n-dimensional code word satisfying the check constraint is generated, specifically including determining an information bit index according to configuration information of the original mold map low density parity check code; the received quaternary sequence is filled into a code word vector with a length of n at a position specified by the information bit index; based on the row echelon form of the check matrix, all check bits of the code word vector are calculated and filled through a back substitution method containing a modulo 4 operation, and an n-dimensional code word satisfying the check constraint is generated, and the process of calculating the check bit value is realized through the following formula: ； wherein, represents the value of the current check bit to be solved, denotes the element of the i-th row and j-th column of the ladder-shaped check matrix, is the other determined bit in the codeword vector except the current check bit to be solved; is the other determined bit in the codeword vector except the current check bit to be solved; denotes modulo 4 operation; performing centralized coding on the n-dimensional code word to map discrete quaternary symbols in the n-dimensional code word into a continuous numerical sequence centered on zero; inputting the continuous numerical sequence into a pre-trained deep decoder network to obtain a reconstructed Gaussian floating point sequence for environment perception of autonomous driving.

2. The high-dimensional lossy source coding method of claim 1, wherein, Before the step of performing centralized coding on the n-dimensional code word, the method further comprises the following steps of: verifying whether the n-dimensional code word satisfies the following check equation: wherein, represents converting an n-dimensional code word from a row vector to a column vector represents a row echelon check matrix; represents a modulo 4 operation.

3. The high-dimensional lossy source coding method of claim 1, wherein, The deep decoder network comprises an input layer, a hidden layer and an output layer, wherein: the input layer maps the continuous numerical sequence to a 4n-dimensional hidden space through linear transformation; the hidden layer adopts a GELU activation function; the output layer maps the 4n-dimensional representation of the hidden layer back to an n-dimensional reconstructed source.

4. The high-dimensional lossy source coding method of claim 3, wherein, The weights of the input layer, the hidden layer and the output layer of the deep decoder network are initialized by Xavier uniform distribution.

5. The high-dimensional lossy source coding method of claim 1, wherein, The centralized coding maps quaternary symbols {0, 1, 2, 3} to {-1.5, -0.5, 0.5, 1.5} respectively.

6. A high-dimensional lossy source coding device for autonomous driving, characterized by, The device is arranged in a vehicle-mounted communication system, and is used for reconstructing point cloud data collected by a sensor, and comprises the following steps of: a quantization unit configured to input n-dimensional source data from a vehicle-mounted sensor into a lossy source encoder network to quantize the n-dimensional source data into a k-dimensional quaternary sequence; an algebraic reconstruction unit configured to perform algebraic reconstruction on the k-dimensional quaternary sequence on an integer ring based on a check matrix of a pre-stored original graph low-density parity-check code to generate an n-dimensional code word satisfying a check constraint; the algebraic reconstruction unit is specifically configured to determine an information bit index according to configuration information of the original graph low-density parity-check code, fill the received quaternary sequence into a position specified by the information bit index in a code word vector with a length of n, and calculate and fill all check bits of the code word vector based on a row echelon form of the check matrix through a back substitution method containing a modulo 4 operation to generate an n-dimensional code word satisfying a check constraint, and the process of calculating the check bit values is implemented through the following formula: wherein, represents the value of the current check bit to be solved, denotes the element of the i-th row and j-th column of the ladder-shaped check matrix, is the other determined bit in the codeword vector except the current check bit to be solved; denotes the modulo 4 operation;​ a centralization unit configured to perform centralized coding on the n-dimensional code word to map discrete quaternary symbols in the n-dimensional code word into a continuous numerical sequence centered on zero; a decoding unit configured to input the continuous numerical sequence into a pre-trained deep decoder network to obtain a reconstructed Gaussian floating point sequence for environment perception of autonomous driving.

7. A high-dimensional lossy source coding device for autonomous driving, characterized in that, The device comprises a processor and a memory; the memory is configured to store instructions; the processor is configured to execute the instructions in the memory to execute the method in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The instructions, when executed on a computer, cause the computer to execute the method in any one of claims 1-5.

Images