Automatic driving perception method and device of vehicle, vehicle and storage medium

Abstract

The application relates to the technical field of intelligent application of automobiles, in particular to an automatic driving perception method and device of a vehicle, a vehicle and a storage medium, wherein the method comprises the following steps: acquiring initial picture data of the surrounding environment of the vehicle; inferring at least one target task according to the initial picture data, and detecting whether preset adversarial samples exist in the initial picture data in the inference process; if the initial picture data does not have the preset adversarial samples, obtaining a perception result of the initial picture data after the inference is completed; otherwise, the preset adversarial samples of the initial picture data are denoised by using a preset denoising strategy, first picture data after denoising is obtained, and the perception result of the initial picture data is obtained again based on the first picture data. Therefore, the problems that all data are usually denoised in the related art, the workload of data processing is large, the perception efficiency is low, and noise of input data of a perception task cannot be effectively avoided are solved.

Description

Technical Field

[0001] This application relates to the field of automotive intelligent application technology, and in particular to an autonomous driving perception method, device, vehicle, and storage medium for a vehicle. Background Technology

[0002] Currently, autonomous driving perception tasks largely employ artificial intelligence algorithms that use neural networks. Research shows that neural networks are not robust to minor perturbations (i.e., adversarial attacks), which poses significant safety risks.

[0003] In related technologies, a denoised image is obtained by denoising the image to be processed; adversarial sample defense is achieved by processing image data based on the denoised image, thereby optimizing the effect of image data processing.

[0004] However, related technologies typically denoise all the data, which is not only labor-intensive and reduces perception efficiency, but also cannot effectively avoid noise problems in the input data of perception tasks. Summary of the Invention

[0005] This application provides an autonomous driving perception method, device, vehicle, and storage medium for vehicles, which solves the problems of related technologies that usually denoise all data, resulting in a large workload for data processing, low perception efficiency, and inability to effectively avoid noise in the input data of perception tasks.

[0006] The first aspect of this application provides a method for autonomous driving perception of a vehicle, comprising the following steps: acquiring initial image data of the vehicle's surrounding environment; inferring at least one target task based on the initial image data, and during the inference process, detecting whether the initial image data contains a preset adversarial sample; if the initial image data does not contain the preset adversarial sample, obtaining a perception result of the initial image data after the inference ends; otherwise, using a preset denoising strategy to denoise the preset adversarial sample in the initial image data to obtain denoised first image data, and re-inferring based on the first image data to obtain a perception result of the initial image data.

[0007] Based on the above technical means, the embodiments of this application perform adversarial sample detection on image data and denoise the image data based on the adversarial samples to achieve accurate denoising. This avoids the huge workload caused by uniform denoising of all data, reduces the workload of denoising processing, and improves processing efficiency while improving perception efficiency. Thus, adversarial sample detection and denoising are combined to systematically solve the noise problem of the input data of perception tasks, effectively strengthen the autonomous driving perception system, and effectively resist the interference of noise on the input data.

[0008] Optionally, in one embodiment of this application, the step of inferring at least one target task based on the initial image data and detecting whether the initial image data contains a preset adversarial sample during the inference process includes: inputting the initial image data into a preset neural network of an autonomous driving perception algorithm to perform inference for at least one target task; during the inference process, using a multi-layer leave-one-out strategy to perform preset adversarial sample detection on multiple intermediate layers of the preset neural network; if a preset adversarial sample exists in any intermediate layer, it is determined that the initial image data contains the preset adversarial sample; otherwise, it is determined that the initial image data does not contain the preset adversarial sample.

[0009] Based on the above technical means, the embodiments of this application can input the initial image data of the vehicle's surrounding environment into multiple intermediate layers in the neural network of the autonomous driving perception algorithm for adversarial sample detection, determine whether the initial image data contains preset adversarial samples, thereby avoiding the problem of unnecessary denoising of data without noise, and realizing the robust reinforcement of the perception algorithm.

[0010] Optionally, in one embodiment of this application, the step of using a preset denoising strategy to denoise the preset adversarial samples of the initial image data to obtain the denoised first image data includes: inputting the image data into a preset high-level guided denoising device to perform preset adversarial sample denoising to obtain the denoised first image data.

[0011] Based on the above technical means, in the embodiments of this application, when there are preset adversarial samples in the initial image data, the image data is input into a high-level guided denoising unit for denoising, thereby systematically solving the problem of noisy input data for the perception task and making the processed image closer to the original image.

[0012] Optionally, in one embodiment of this application, before re-inferring the perception result of the initial image data based on the first image data, the method further includes: continuing to detect whether the preset adversarial sample exists in the first image data during the inference process; if the preset adversarial sample does not exist in the first image data, then the perception result of the initial image data is obtained after the inference ends; otherwise, the preset denoising strategy is used to denoise the preset adversarial sample in the first image data to obtain the denoised second image data, until the preset adversarial sample does not exist or the number of inferences reaches a preset number, then the inference ends and the perception result of the initial image data is obtained.

[0013] Based on the above technical means, the embodiments of this application can perform adversarial sample detection on the obtained first image data, and repeat the steps of adversarial sample detection, adversarial sample denoising, and inference again until there are no adversarial samples or the inference reaches a certain number of times, then end the inference and obtain the perception result of the initial image data, thereby meeting the reinforcement requirements of the visual perception task, improving the accuracy of image denoising, and making the image data processing effect reach the best.

[0014] A second aspect of this application provides an autonomous driving perception device for a vehicle, comprising: an acquisition module for acquiring initial image data of the vehicle's surrounding environment; an inference module for inferring at least one target task based on the initial image data, and detecting whether a preset adversarial sample exists in the initial image data during the inference process; and a perception module for obtaining a perception result of the initial image data after the inference ends if the preset adversarial sample does not exist in the initial image data, otherwise denoising the preset adversarial sample in the initial image data using a preset denoising strategy to obtain denoised first image data, and re-inferring the perception result of the initial image data based on the first image data.

[0015] Optionally, in one embodiment of this application, the inference module is further configured to input the initial image data into a preset neural network of an autonomous driving perception algorithm to perform inference for at least one target task. During the inference process, a multi-layer leave-one-out strategy is used to perform preset adversarial sample detection on multiple intermediate layers of the preset neural network. If a preset adversarial sample exists in any intermediate layer, it is determined that the initial image data contains the preset adversarial sample; otherwise, it is determined that the initial image data does not contain the preset adversarial sample.

[0016] Optionally, in one embodiment of this application, the perception module is further configured to input the image data into a preset high-level guided denoising device to perform preset adversarial sample denoising, thereby obtaining the denoised first image data.

[0017] Optionally, in one embodiment of this application, before re-inferring the perception result of the initial image data based on the first image data, the method further includes: a detection module, configured to continue detecting whether the preset adversarial sample exists in the first image data during the inference process; and a processing module, configured to obtain the perception result of the initial image data after the inference ends if the preset adversarial sample does not exist in the first image data, otherwise, to denoise the preset adversarial sample of the first image data using a preset denoising strategy to obtain denoised second image data, until the preset adversarial sample does not exist or the number of inferences reaches a preset number, at which point the inference ends and the perception result of the initial image data is obtained.

[0018] A third aspect of this application provides a vehicle, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the autonomous driving perception method for the vehicle as described in the above embodiments.

[0019] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the autonomous driving perception method for a vehicle as described in the above embodiments.

[0020] Therefore, this application has at least the following beneficial effects:

[0021] 1. By performing adversarial example detection on image data, and then denoising the image data based on the adversarial examples, precise denoising is achieved. This avoids the huge workload caused by uniform denoising of all data, reduces the workload of denoising processing, and improves processing efficiency while also improving perception efficiency. Thus, by combining adversarial example detection with denoising, the noise problem of the input data of perception tasks is systematically solved, effectively strengthening the autonomous driving perception system and effectively resisting the interference of noise on the input data.

[0022] 2. Initial image data of the vehicle's surrounding environment can be input into multiple intermediate layers of the neural network of the autonomous driving perception algorithm for adversarial sample detection. This determines whether there are pre-set adversarial samples in the initial image data, thereby avoiding the problem of unnecessary denoising of data without noise and achieving robust reinforcement of the perception algorithm.

[0023] 3. When the initial image data contains pre-set adversarial samples, the image data is input into a high-level guided denoising unit for denoising, thereby systematically solving the problem of noisy input data for the perception task and making the processed image closer to the original image.

[0024] 4. Adversarial sample detection can be performed on the obtained first image data, and the steps of adversarial sample detection, adversarial sample denoising, and inference can be repeated until no adversarial samples are found or the inference reaches a certain number of times. Then, the inference ends and the perception result of the initial image data is obtained, thereby meeting the reinforcement requirements of the visual perception task, improving the accuracy of image denoising, and making the image data processing effect optimal.

[0025] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0026] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0027] Figure 1 This is a flowchart of an autonomous driving perception method for a vehicle provided according to an embodiment of this application;

[0028] Figure 2 This is a schematic diagram of an autonomous driving perception method for a vehicle according to an embodiment of this application;

[0029] Figure 3 This is a block diagram of an autonomous driving perception device for a vehicle according to an embodiment of this application;

[0030] Figure 4 This is a structural schematic diagram of a vehicle provided according to an embodiment of this application.

[0031] Explanation of reference numerals in the attached diagram: Acquisition module-100, Inference module-200, Perception module-300, Memory-401, Processor-402, Communication interface-403. Detailed Implementation

[0032] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0033] The following description, with reference to the accompanying drawings, describes an autonomous driving perception method, apparatus, vehicle, and storage medium for vehicles according to embodiments of this application. Addressing the problems mentioned in the background section, this application provides an autonomous driving perception method for vehicles. In this method, initial image data of the vehicle's surrounding environment is acquired; at least one target task is inferred based on the initial image data; during the inference process, the presence of pre-set adversarial samples in the initial image data is detected; if no pre-set adversarial samples are found, the perception result of the initial image data is obtained after the inference ends; otherwise, a pre-set denoising strategy is used to denoise the pre-set adversarial samples in the initial image data to obtain denoised first image data, and the perception result of the initial image data is obtained again based on the first image data. Therefore, this application combines adversarial sample detection with denoising, avoiding unnecessary denoising of noise-free data, thereby achieving robust reinforcement of the perception algorithm. This solves the problems of related technologies, which typically denoise all data, resulting in a large workload for data processing, low perception efficiency, and an inability to effectively avoid noise in the input data of the perception task.

[0034] Specifically, Figure 1 This is a flowchart illustrating an autonomous driving perception method for a vehicle provided in an embodiment of this application.

[0035] like Figure 1As shown, the autonomous driving perception method for this vehicle includes the following steps:

[0036] In step S101, initial image data of the vehicle's surrounding environment is acquired.

[0037] This embodiment combines adversarial example detection and denoising of image data. First, a multi-layer leave-one-out method is used to detect adversarial examples. The detected adversarial examples are then input into a high-level guided denoising unit for further denoising, thus systematically solving the problem of noisy input data for perception tasks. Therefore, before detecting adversarial examples in image data, this embodiment can first acquire initial image data of the vehicle's surrounding environment. In actual execution, this embodiment can use an in-vehicle camera to capture images of the vehicle's surrounding environment, thereby obtaining initial image data of the vehicle's surrounding environment.

[0038] In step S102, at least one target task is inferred based on the initial image data, and during the inference process, it is detected whether there are pre-set adversarial samples in the initial image data.

[0039] This application embodiment can input image data into the neural network of a perception algorithm for inference to obtain the target task of the autonomous vehicle. During the inference process, it detects whether the initial image data contains samples that may mislead the autonomous vehicle in making accurate classifications and interfere with driving decisions. The target task includes, but is not limited to, autonomous driving perception tasks such as object detection, lane detection, and semantic segmentation.

[0040] In one embodiment of this application, inferring at least one target task based on initial image data, and detecting whether the initial image data contains pre-set adversarial samples during the inference process, includes: inputting the initial image data into a pre-set neural network of an autonomous driving perception algorithm to infer at least one target task; during the inference process, using a multi-layer leave-one-out strategy to perform pre-set adversarial sample detection on multiple intermediate layers of the pre-set neural network, and determining that the initial image data contains pre-set adversarial samples when a pre-set adversarial sample exists in any intermediate layer, otherwise determining that the initial image data does not contain pre-set adversarial samples.

[0041] It is understood that, in this embodiment of the application, initial image data of the vehicle's surrounding environment can be input into the neural network of the autonomous driving perception algorithm, and a multi-layer leave-one-out strategy can be used to perform adversarial example detection on multiple intermediate layers of the neural network to determine whether adversarial examples exist in the initial image data. The specific process is as follows:

[0042] 1) First, record the difference between the left-one value and the original value of a single node.

[0043]

[0044] It refers to the difference between the retained value and the original value;

[0045] To retain one value;

[0046] The original value;

[0047] 2) Calculate the interquartile range

[0048]

[0049] in,

[0050] Since the interquartile range of data containing adversarial examples is relatively large, this application embodiment can determine whether there are adversarial examples in the initial image based on the magnitude of the interquartile range of the adversarial example data.

[0051] In step S103, if the initial image data does not contain any preset adversarial samples, the perception result of the initial image data is obtained after the inference ends; otherwise, the preset denoising strategy is used to denoise the preset adversarial samples of the initial image data to obtain the denoised first image data, and the perception result of the initial image data is obtained by re-inference based on the first image data.

[0052] It is understood that the embodiments of this application can determine whether to perform adversarial sample denoising based on whether there are preset adversarial samples in the initial image data. If no adversarial samples are detected, the entire inference process is completed directly to obtain the result of the perception task. If preset adversarial samples are detected, denoising is performed according to a preset denoising strategy, and the denoised image data is used for inference to obtain the perception result. Thus, adversarial sample detection and denoising are combined to systematically solve the noise problem of the input data of the perception task, thereby strengthening the autonomous driving perception system and resisting the interference of noise on the input data.

[0053] In one embodiment of this application, a preset adversarial sample of the initial image data is denoised using a preset denoising strategy to obtain the denoised first image data, including: inputting the image data into a preset high-level guided denoiser to perform preset adversarial sample denoising to obtain the denoised first image data.

[0054] It is understood that when the initial image data contains pre-defined adversarial examples, this embodiment of the application can input the image data into a high-level guided denoising unit for denoising, thereby systematically solving the problem of noisy input data for the perception task and making the processed image closer to the original image. The pre-defined high-level guided denoising unit can be a feature-guided denoising unit, a logarithmic guided denoising unit, etc., and those skilled in the art can choose according to the actual situation without specific limitations. In this embodiment of the application, a feature-guided denoising unit or a logarithmic guided denoising unit is used as an example to achieve effective denoising of adversarial examples and meet the user's denoising needs.

[0055] In one embodiment of this application, before obtaining the perception result of the initial image data based on the first image data through re-inference, the method further includes: continuing to detect whether there are preset adversarial samples in the first image data during the inference process; if there are no preset adversarial samples in the first image data, then obtaining the perception result of the initial image data after the inference ends; otherwise, using a preset denoising strategy to denoise the preset adversarial samples in the first image data to obtain denoised second image data, until there are no preset adversarial samples or the number of inferences reaches a preset number, then ending the inference and obtaining the perception result of the initial image data.

[0056] It is understood that, in this embodiment, after obtaining the initial image data containing adversarial examples and having undergone denoising, the denoised image data can be input again into the neural network for the perception task. To improve efficiency, multi-layer leave-one-out adversarial example detection can be omitted, and the perception result of the initial image data can be obtained after inference. However, given the high robustness requirements of the perception task and the redundancy of computational resources, this embodiment can re-perform adversarial example detection on the denoised image data and repeat the steps of adversarial example detection, adversarial example denoising, and inference again until no adversarial examples exist or the inference reaches a certain number of iterations. Then, the inference ends, and the perception result of the initial image data is obtained, thereby meeting the reinforcement requirements of the visual perception task, improving the accuracy of image denoising, and making the processed image closer to the original image.

[0057] The following will provide a detailed description of the vehicle autonomous driving perception method of this application embodiment through a specific implementation process, such as... Figure 2 As shown, it includes the following steps:

[0058] Step 1) Input the image data into the neural network of the perception algorithm for inference;

[0059] Step 2) Use the leave-one-out-of-multilayer method to perform adversarial sample detection on multiple intermediate layers of the neural network to determine whether there are adversarial samples in the image;

[0060] Step 3) If no adversarial example is detected in Step 2), the entire inference process is completed directly, and the result of the perception task is obtained. If an adversarial example is detected in Step 2), the corresponding original image is input into the high-level guided denoising unit for denoising.

[0061] Step 4) The denoised image is re-inputted into the neural network for inference to obtain the result of the perception task.

[0062] The autonomous driving perception method for vehicles proposed in this application involves acquiring initial image data of the vehicle's surrounding environment; inferring at least one target task based on the initial image data; and detecting whether pre-set adversarial samples exist in the initial image data during the inference process. If no pre-set adversarial samples exist, the perception result of the initial image data is obtained after the inference ends; otherwise, a pre-set denoising strategy is used to denoise the pre-set adversarial samples in the initial image data to obtain denoised first image data, and the perception result of the initial image data is obtained again based on the first image data. Therefore, this application combines adversarial sample detection with denoising, systematically solving the problem of noisy input data for perception tasks, thereby strengthening the autonomous driving perception system. It addresses the problems of related technologies that typically denoise all data, resulting in a large workload for data processing, low perception efficiency, and an inability to effectively avoid noise in the input data of perception tasks.

[0063] Next, referring to the accompanying drawings, an autonomous driving perception device for a vehicle according to an embodiment of this application is described.

[0064] Figure 3 This is a block diagram of an autonomous driving perception device for a vehicle according to an embodiment of this application.

[0065] like Figure 3 As shown, the autonomous driving perception device 10 of the vehicle includes: an acquisition module 100, an inference module 200, and a perception module 300.

[0066] The system includes an acquisition module 100 for acquiring initial image data of the vehicle's surrounding environment; an inference module 200 for inferring at least one target task based on the initial image data, and detecting whether there are pre-set adversarial samples in the initial image data during the inference process; and a perception module 300 for obtaining the perception result of the initial image data after the inference ends if there are no pre-set adversarial samples in the initial image data, otherwise using a pre-set denoising strategy to denoise the pre-set adversarial samples in the initial image data to obtain denoised first image data, and re-inferring the perception result of the initial image data based on the first image data.

[0067] In one embodiment of this application, the inference module 200 is further configured to input the initial image data into a preset neural network of the autonomous driving perception algorithm to perform inference for at least one target task. During the inference process, a multi-layer leave-one-out strategy is used to perform preset adversarial sample detection on multiple intermediate layers of the preset neural network. If a preset adversarial sample exists in any intermediate layer, it is determined that the initial image data contains a preset adversarial sample; otherwise, it is determined that the initial image data does not contain a preset adversarial sample.

[0068] In one embodiment of this application, the perception module 300 is further configured to input image data into a preset high-level guided denoising device to perform preset adversarial sample denoising, thereby obtaining the denoised first image data.

[0069] In one embodiment of this application, before obtaining the perception result of the initial image data based on the first image data through re-inference, the method further includes: a detection module, used to continue detecting whether there are preset adversarial samples in the first image data during the inference process; and a processing module, used to obtain the perception result of the initial image data after the inference ends if there are no preset adversarial samples in the first image data, otherwise to denoise the preset adversarial samples in the first image data using a preset denoising strategy to obtain denoised second image data, until there are no preset adversarial samples or the number of inferences reaches a preset number, at which point the inference ends and the perception result of the initial image data is obtained.

[0070] It should be noted that the foregoing explanation of the vehicle autonomous driving perception method embodiment also applies to the vehicle autonomous driving perception device of this embodiment, and will not be repeated here.

[0071] The autonomous driving perception device for vehicles proposed in this application acquires initial image data of the vehicle's surrounding environment; infers at least one target task based on the initial image data; and during the inference process, detects whether there are pre-set adversarial samples in the initial image data; if no pre-set adversarial samples are found, the perception result of the initial image data is obtained after the inference ends; otherwise, a pre-set denoising strategy is used to denoise the pre-set adversarial samples in the initial image data to obtain denoised first image data, and the perception result of the initial image data is obtained again based on the first image data. Therefore, this application combines adversarial sample detection with denoising, avoiding the problem of unnecessary denoising of data without noise, thereby achieving robust reinforcement of the perception algorithm. It solves the problems of related technologies that typically denoise all data, resulting in a large workload for data processing, low perception efficiency, and an inability to effectively avoid noise in the input data of the perception task.

[0072] Figure 4 A schematic diagram of the structure of a vehicle provided in an embodiment of this application. The vehicle may include:

[0073] The memory 401, the processor 402, and the computer program stored on the memory 401 and capable of running on the processor 402.

[0074] When the processor 402 executes the program, it implements the vehicle autonomous driving perception method provided in the above embodiments.

[0075] Furthermore, the vehicle also includes:

[0076] Communication interface 403 is used for communication between memory 401 and processor 402.

[0077] The memory 401 is used to store computer programs that can run on the processor 402.

[0078] The memory 401 may include high-speed RAM (Random Access Memory) memory, and may also include non-volatile memory, such as at least one disk storage.

[0079] If the memory 401, processor 402, and communication interface 403 are implemented independently, then the communication interface 403, memory 401, and processor 402 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0080] Optionally, in a specific implementation, if the memory 401, processor 402, and communication interface 403 are integrated on a single chip, then the memory 401, processor 402, and communication interface 403 can communicate with each other through an internal interface.

[0081] Processor 402 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of this application.

[0082] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described autonomous driving perception method for vehicles.

[0083] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0084] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0085] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0086] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0087] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0088] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for autonomous driving perception of a vehicle, characterized in that, Includes the following steps: Acquire initial image data of the vehicle's surrounding environment; Infer at least one target task based on the initial image data, and during the inference process, detect whether the initial image data contains pre-set adversarial samples; If the initial image data does not contain the preset adversarial sample, the perception result of the initial image data is obtained after the inference ends; otherwise, the preset denoising strategy is used to denoise the preset adversarial sample of the initial image data to obtain the denoised first image data, and the perception result of the initial image data is obtained by re-inference based on the first image data. The step of inferring at least one target task based on the initial image data, and detecting whether the initial image data contains pre-set adversarial samples during the inference process, includes: The initial image data is input into a preset neural network of the autonomous driving perception algorithm to perform inference for at least one target task; During the inference process, a multi-layer leave-one-out strategy is used to perform pre-defined adversarial sample detection on multiple intermediate layers of the pre-defined neural network. If a pre-defined adversarial sample exists in any intermediate layer, it is determined that the initial image data contains the pre-defined adversarial sample; otherwise, it is determined that the initial image data does not contain the pre-defined adversarial sample. The step of denoising the initial image data using a preset denoising strategy to obtain the denoised first image data includes: The image data is input into a preset high-level guided denoising device to perform preset adversarial sample denoising, and the denoised first image data is obtained. The preset high-level guided denoising device is a feature-guided denoising device and a logarithmic guided denoising device. The specific process of the multi-level leave-one-out strategy is as follows: First, record the difference between the left-one value and the original value of a single node. It refers to the difference between the retained value and the original value; To retain one value; The original value; Calculate the interquartile range 。 2. The method according to claim 1, characterized in that, Before obtaining the perception result of the initial image data through re-inference based on the first image data, the process further includes: During the inference process, the presence of the preset adversarial sample in the first image data is continuously detected. If the first image data does not contain the preset adversarial sample, the perception result of the initial image data is obtained after the inference ends. Otherwise, the preset denoising strategy is used to denoise the preset adversarial sample of the first image data to obtain the denoised second image data. The inference ends when the preset adversarial sample does not exist or the number of inferences reaches a preset number, and the perception result of the initial image data is obtained.

3. An autonomous driving perception device for a vehicle, used to perform the method as described in any one of claims 1-2; characterized in that, include: The acquisition module is used to acquire initial image data of the vehicle's surrounding environment; The inference module is used to infer at least one target task based on the initial image data, and during the inference process, detect whether there are preset adversarial samples in the initial image data; The perception module is configured to obtain the perception result of the initial image data after the inference ends if the preset adversarial sample does not exist in the initial image data; otherwise, it uses a preset denoising strategy to denoise the preset adversarial sample in the initial image data to obtain denoised first image data, and then re-infers the perception result of the initial image data based on the first image data.

4. The apparatus according to claim 3, characterized in that, The inference module is further used for: The initial image data is input into a preset neural network of the autonomous driving perception algorithm to perform inference for at least one target task; During the inference process, a multi-layer leave-one-out strategy is used to perform pre-defined adversarial sample detection on multiple intermediate layers of the pre-defined neural network. If a pre-defined adversarial sample exists in any intermediate layer, it is determined that the initial image data contains the pre-defined adversarial sample; otherwise, it is determined that the initial image data does not contain the pre-defined adversarial sample.

5. The apparatus according to claim 3, characterized in that, The sensing module is further used for: The image data is input into a preset high-level guided denoising device to perform preset adversarial sample denoising, resulting in the denoised first image data.

6. The apparatus according to any one of claims 3-5, characterized in that, Before obtaining the perception result of the initial image data through re-inference based on the first image data, the process further includes: The detection module is used to continue detecting whether the preset adversarial sample exists in the first image data during the inference process; The processing module is configured to, if the first image data does not contain the preset adversarial sample, obtain the perception result of the initial image data after the inference ends; otherwise, it uses a preset denoising strategy to denoise the preset adversarial sample of the first image data to obtain the denoised second image data, until the preset adversarial sample does not exist or the number of inferences reaches a preset number, then end the inference and obtain the perception result of the initial image data.

7. A vehicle, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the autonomous driving perception method for a vehicle as described in any one of claims 1-2.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the autonomous driving perception method for the vehicle as described in any one of claims 1-2.

Images