Key privacy information desensitization driving record method and system and storage medium

Abstract

The application discloses a kind of key privacy information desensitization driving record method, system and storage medium, the method includes: obtaining the data set of driving record image;Key privacy information positioning model is obtained by training network model architecture using data set, the positioning of key privacy information in driving record image is determined using model, and the desensitization area to be obtained;Desensitization area to be input into privacy information desensitization module based on chaotic encryption system, generate encrypted image, realize the desensitization of desensitization area to be;The image in the actual vehicle driving process collected is input into key privacy information positioning model, the desensitization area to be output is input into privacy information desensitization module, the information of the original image corresponding position is covered by the encrypted image generated, and the image after desensitization is saved in storage medium.The network model architecture and privacy information desensitization module are constructed in the application, the desensitization of key privacy information in driving record is realized, and the key information of driving record is retained.

Description

Technical Field

[0001] This invention belongs to the field of image processing and deep learning technology, and relates to a method, system, terminal device and storage medium for desensitizing key privacy information in vehicle recording. Background Technology

[0002] To regulate automotive data processing activities, protect the legitimate rights and interests of individuals and organizations, safeguard national security and public interests, and promote the rational development and utilization of automotive data, the State has formulated the "Several Provisions on the Management of Automotive Data Security (Trial Implementation)," which came into effect on October 1, 2021. According to Article 6, Point 4 of the provisions, the State encourages the lawful, rational, and effective use of automotive data and advocates that automotive data processors adhere to the principle of de-identification in their automotive data processing activities, and perform anonymization and de-identification as much as possible.

[0003] The dashcam recording system can record information such as time, speed, video, and sound during a vehicle's journey. This not only protects the driver's legal rights and supervises the driver's legal driving, but also provides effective and accurate evidence for traffic police, courts, insurance companies, and other institutions when handling traffic accident cases, facilitating their quick and accurate processing of cases.

[0004] In recent years, with the continuous development of dashcam technology, different dashcams are suitable for different scenarios, such as high-definition, night vision, wide-angle, and dual-lens models, with their recording range and capabilities constantly expanding and improving. However, due to the randomness and wide range of vehicle driving routes, the continuous recording by dashcams inevitably leads to the recording of a large amount of personal privacy information and critical information involving public safety. Since the information recorded by dashcams is stored in the data storage medium on the device itself, it can be easily accessed by individuals. The data security of critical privacy information such as facial and license plate information cannot be guaranteed. Therefore, before storing the data on the storage medium, it is necessary to perform anonymization processing on the dashcam data.

[0005] Unlike cameras that are limited to specific locations and angles, dashcams are affected by various factors during actual operation, such as license plates and facial images that are offset at large angles, or facial images of people wearing masks. Current dashcam recording methods are prone to problems in accurately locating the positions of faces and license plates, and cannot completely desensitize the data. Summary of the Invention

[0006] In view of this, the present invention provides a method, system, terminal device, and storage medium for desensitizing key privacy information in vehicle recordings. By constructing a network model architecture, the location of key privacy information in the vehicle recording image is determined, solving the problems of numerous interferences in the recorded images during driving and the easy omission or false detection of key privacy information. By constructing a privacy information desensitization module and using a chaotic encryption algorithm to desensitize the located key privacy information, it ensures that individuals cannot obtain key privacy information through the storage medium, solving the problem of desensitizing key privacy information in vehicle recordings, while retaining the key information recording function of the vehicle recording system.

[0007] The first objective of this invention is to provide a method for desensitizing key privacy information in a vehicle recording system.

[0008] The second objective of this invention is to provide a vehicle recording system that desensitizes key privacy information.

[0009] A third objective of this invention is to provide a terminal device.

[0010] A fourth objective of this invention is to provide a storage medium.

[0011] The first objective of this invention can be achieved by adopting the following technical solution:

[0012] A method for de-identifying key privacy information in a vehicle recording system, the method comprising:

[0013] Acquire a dataset of dashcam images; the dashcam images include at least one of a human eye and a license plate;

[0014] A network model architecture is constructed to identify human eye and license plate information in dashcam images. The network model architecture is trained using the dataset to obtain a key privacy information localization model. The key privacy information localization model is used to determine the location of key privacy information in dashcam images, thereby obtaining the region to be desensitized. The key privacy information includes face and license plate.

[0015] The area to be desensitized is input into the privacy information desensitization module based on the chaotic encryption system to generate an encrypted image, thereby achieving the desensitization of the area to be desensitized.

[0016] The images collected during the actual vehicle driving process are input into the key privacy information positioning model. The output area to be desensitized is input into the privacy information desensitization module. The generated encrypted image overwrites the information at the corresponding position of the original image to obtain the desensitized image. The desensitized image is then stored in a storage medium.

[0017] Furthermore, the step of using the key privacy information location model to determine the location of key privacy information in the dashcam image and obtaining the region to be desensitized includes:

[0018] A key privacy information localization model was used to locate human eyes and license plates in dashcam images;

[0019] The location of the face is determined based on the positioning of the human eye;

[0020] The location information of the face and license plate is combined to obtain the area to be desensitized.

[0021] Furthermore, the network model architecture includes a preprocessing module, a feature extraction skeleton module, a feature fusion pyramid module, and output Loss and NMS modules, wherein:

[0022] The preprocessing module is used to preprocess the vehicle recording images;

[0023] The skeleton module for feature extraction uses the Focus structure and the CSP structure in CSPNet to extract features from the preprocessed image; wherein the Focus structure is used to adjust the size of the feature map, and the CSP structure includes convolutional layers and residual units to extract features from the input feature map.

[0024] The feature fusion pyramid module adopts a fusion structure of FPN and PAN. The FPN structure is an upsampling process that conveys strong semantic information from top to bottom; the PAN structure is a feature pyramid from bottom to top that conveys strong localization information from bottom to top. The pyramid module integrates parameters and extracts features from feature layers of different sizes from different backbone layers, thereby extracting more information based on the extracted features.

[0025] The Loss uses GIOU_Loss as the loss function for the target detection box, and the NMS module is used to filter the predicted target detection boxes.

[0026] Furthermore, a key privacy information localization model is used to locate human eyes in dashcam images, including:

[0027] The NMS module is used to filter the obtained predicted target detection boxes, and the filtered predicted target detection boxes are used as the predicted boxes for human eyes to achieve human eye localization.

[0028] Similarly, the selected predicted target detection boxes are used as the predicted boxes for license plates to achieve license plate localization.

[0029] Furthermore, let intersection A be the area of ​​the overlapping portion of the ground truth detection boxes and the predicted detection boxes, and union B be the total area occupied by the ground truth detection boxes and the predicted detection boxes, then:

[0030]

[0031] Let C be the area of ​​the minimum bounding rectangle of the ground truth bounding box and the predicted bounding box. Then we get the loss function GIOU_Loss:

[0032]

[0033] Furthermore, the preprocessing includes Mosaic data augmentation, cmBN, and adaptive image scaling, wherein the Mosaic data augmentation includes random arrangement, random scaling, and random cropping of the image.

[0034] Furthermore, the area to be desensitized is used as the image to be encrypted;

[0035] The step of inputting the region to be desensitized into the privacy information desensitization module based on the chaotic encryption system to generate an encrypted image includes:

[0036] An encryption sequence is generated based on the chaotic mapping function and the pixel size of the image to be encrypted;

[0037] Based on the encrypted sequence, a chaotic sequence is obtained;

[0038] Generate a sequence of information to be encrypted based on the image to be encrypted;

[0039] The chaotic sequence is used to encrypt the information to be encrypted, and an encrypted image is generated based on the encrypted sequence. The second objective of this invention can be achieved by adopting the following technical solution:

[0040] A vehicle recording system for de-identifying key privacy information, the system comprising:

[0041] A dataset acquisition module is used to acquire a dataset of driving recorder images; the driving recorder images include at least one of a human eye and a license plate;

[0042] A key privacy information localization module is used to build a network model architecture to identify human eye and license plate information in dashcam images; the network model architecture is trained using the dataset to obtain a key privacy information localization model; the key privacy information localization model is used to determine the location of key privacy information in dashcam images to obtain the region to be desensitized; the key privacy information includes face and license plate.

[0043] The key privacy information desensitization module is used to input the area to be desensitized into the privacy information desensitization module based on the chaotic encryption system, generate an encrypted image, and achieve desensitization of the area to be desensitized;

[0044] The desensitized image generation module is used to input the collected images of the actual vehicle driving process into the key privacy information positioning model, input the output area to be desensitized into the privacy information desensitization module, and overwrite the information at the corresponding position of the original image with the generated encrypted image to obtain the desensitized image; the desensitized image is then stored in the storage medium.

[0045] The third objective of this invention can be achieved by adopting the following technical solution:

[0046] A terminal device includes a processor and a memory for storing a processor-executable program. When the processor executes the program stored in the memory, it implements the above-mentioned method for desensitizing key privacy information during vehicle recording.

[0047] The fourth objective of this invention can be achieved by adopting the following technical solution:

[0048] A storage medium storing a program that, when executed by a processor, implements the aforementioned method for desensitizing key privacy information in vehicle recording.

[0049] The present invention has the following advantages over the prior art:

[0050] 1. This invention utilizes a network model architecture to locate critical privacy information in dashcam images. It addresses issues such as masked faces, large-angle profile views, and license plates from new energy vehicles, resolving problems of numerous image interferences and the potential for missed or false detections of critical privacy information during driving.

[0051] 2. This invention constructs a privacy information desensitization module and employs a chaotic encryption algorithm to desensitize key privacy information located by the key privacy information location model. This ensures that individuals cannot access key privacy information through storage media, but law enforcement agencies such as public security can obtain the corresponding key privacy information by decrypting the image when necessary. This solves the problem of desensitizing key privacy information while retaining the key information recording function of the dashcam. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0053] Figure 1 This is a flowchart of a driving recorder method for desensitizing key privacy information in Embodiment 1 of the present invention.

[0054] Figure 2This is a flowchart of the key privacy information location module in Embodiment 1 of the present invention.

[0055] Figure 3 This is a flowchart of the key privacy information desensitization module in Embodiment 1 of the present invention.

[0056] Figure 4 This is a structural block diagram of a vehicle recording system for desensitizing key privacy information, as shown in Embodiment 2 of the present invention.

[0057] Figure 5 This is a structural block diagram of the terminal device according to Embodiment 3 of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be understood that the specific embodiments described are merely used to explain this application and are not intended to limit this application.

[0059] Example 1:

[0060] This embodiment is developed using the PyCharm development environment, the OpenCV computer vision library, and the PyTorch deep learning framework. OpenCV encompasses a large number of encapsulated function interfaces related to image processing, enabling the completion of relevant image processing tasks. The PyTorch deep learning framework is a Python-first deep learning framework that not only achieves powerful GPU acceleration but also supports dynamic neural networks. The PyCharm development environment on the Windows platform is currently one of the preferred development environments for completing image processing and machine learning tasks.

[0061] like Figure 1 As shown in the figure, this embodiment provides a method for de-identifying key privacy information in a vehicle recording system, including:

[0062] S101. Obtain the dataset of driving recorder images.

[0063] This dataset contains images of human eyes and license plates captured using high-definition cameras and other equipment under a pre-defined environment. The dataset includes images of human eye feature points, regular license plates, license plates at large angles, and license plates of new energy vehicles. The pre-defined environment refers to an environment that matches road conditions and can basically reflect real-world road conditions. The dataset is randomly selected in an 8:2 ratio to obtain the training and test sets.

[0064] In this embodiment, the dataset contains 10,000 driving record images of human eyes and license plates. The training set contains 4,000 images of both human eyes and license plates, 2,000 images of only human eyes, and 2,000 images of only license plates. The test set contains 1,000 images of both human eyes and license plates, 500 images of only human eyes, and 500 images of only license plates.

[0065] S102. Build a network model architecture to identify human eye and license plate information in dashcam images; train the network model architecture using the dataset to obtain a key privacy information localization model; use the key privacy information localization model to determine the location of key privacy information in dashcam images and obtain the region to be desensitized.

[0066] A network model architecture capable of recognizing features related to human eyes and license plates was constructed. The relevant parameters of human eyes and license plates were trained using training and testing sets to obtain a key privacy information localization model. This model is used to detect the positions of human eyes and license plates, and further determine the position of the face by using the position of human eyes, thereby avoiding errors or large errors in face position recognition caused by wearing masks. At the same time, the license plate position is combined to complete the localization of the entire key privacy information.

[0067] (1) Build the network model architecture.

[0068] Utilize deep learning technology to build network model architecture.

[0069] The network model architecture includes a preprocessing module, a skeleton module for feature extraction, a pyramid module for feature fusion, and loss and NMS modules at the output end. The output of each module is the input of the next module.

[0070] (1-1) Preprocessing module.

[0071] The preprocessing module processes the original input images to highlight their features, enrich the dataset, and reduce the training requirements of the GPU.

[0072] Preprocessing includes methods such as Mosaic data augmentation, cmBN, and adaptive image scaling. Mosaic data augmentation stitches together images in the dataset by randomly arranging, scaling, and cropping them. Adaptive image scaling uniformly scales the input images to a standard size to facilitate subsequent network processing, making it an effective image preprocessing method.

[0073] In this embodiment, the preprocessing module uses methods such as Mosaic data augmentation and adaptive image scaling to process images. The Mosaic data augmentation method includes random arrangement, random scaling, and random cropping to stitch images in the dataset. Adaptive image scaling scales the input images to the same standard size, making the images easier for subsequent modules to process.

[0074] (1-2) Skeleton module for feature extraction.

[0075] The skeleton module for feature extraction uses a network structure to extract image features, employing the Focus structure and the CSP structure in CSPNet.

[0076] The Focus structure performs a series of slicing operations on the input image, followed by a convolution operation with 32 convolutional kernels, thereby adjusting the size of the feature map.

[0077] In this embodiment, the original input image of size 608×608×3 is sliced ​​by the Focus structure and transformed into a feature map of size 304×304×12. Then, after convolution with 32 convolution kernels, it is transformed into a feature map of size 304×304×32.

[0078] The CSP structure consists of convolutional layers and residual units, used to extract features from the input feature map. Each CSP module has a 3×3 kernel with a stride of 2, serving as a downsampling mechanism. After passing through five CSP modules, the input feature map is downsampled five times, resulting in a 19×19 feature map.

[0079] (1-3) Pyramid module for feature fusion.

[0080] The feature fusion pyramid module is implemented using a fusion structure of FPN and PAN. The FPN structure is an upsampling process that conveys strong semantic information from top to bottom, upsampling the feature map size from 19×19 to 76×76. The PAN structure is a bottom-up feature pyramid that conveys strong localization information, downsampling the feature map size from 76×76 to 19×19, corresponding to the minimum and maximum anchor boxes, respectively. The two pyramid models integrate parameters and extract features from feature layers of different sizes from different backbone layers, which can extract more information from the feature maps.

[0081] (1-4) Loss and NMS modules at the output end.

[0082] The output modules Loss and NMS are used to select the loss function and filter the predicted bounding boxes for object detection tasks. GIOU_Loss is used as the loss for object detection boxes, which is an improvement and supplement to the IOU loss function. Let the intersection A be the area of ​​the overlapping portion of the ground truth bounding box and the predicted bounding box, and the union B be the total area occupied by the two boxes. Then:

[0083]

[0084] Let C be the area of ​​the minimum bounding rectangle of the two boxes. Then the loss function can be expressed as follows:

[0085]

[0086] Finally, the NMS method is used to filter the predicted boxes, removing redundant, overlapping, and inaccurate predicted boxes.

[0087] (2) Use the dataset to train the network model architecture to obtain the key privacy information location model, use the key privacy information location model to determine the location of key privacy information in the driving record image, and obtain the area to be desensitized.

[0088] The relevant parameters of human eyes and license plates in the key privacy information localization model are trained using the training set and test set of human eyes and license plates, resulting in a well-trained key privacy information localization model.

[0089] In this embodiment, the Adam algorithm is used as the optimization parameter of the model. The training period is set to 300, the batch size is set to 128, and the learning rate is 0.0001. The training of the model is completed so that the model converges and the localization of human eyes and license plates is achieved.

[0090] By identifying the coordinates of the human eyes, the face is located and labeled according to the usual proportions of the eyes and face, determining the position coordinates of the face and labeling them as a predicted bounding box; the combination of the predicted bounding boxes of the face and the license plate is then output. The flowchart of the key privacy information location model is as follows: Figure 2 As shown.

[0091] S103. Input the collected images of the actual vehicle driving process into the key privacy information positioning model, and input the output area to be desensitized into the privacy information desensitization module to achieve desensitization of the area to be desensitized; based on the desensitized area to be desensitized, obtain the desensitized image.

[0092] The critical privacy information location module outputs the region to be de-identified (i.e., the predicted bounding box), with each predicted bounding box representing one image to be de-identified. The critical privacy information de-identification module is used to perform de-identification processing on the images to be de-identified.

[0093] like Figure 3As shown, a chaotic encryption algorithm is used to de-identify critical privacy information (i.e., the area to be de-identified). The chaotic encryption algorithm requires some known parameters as encryption seed parameters. These parameters are randomly generated and uniformly stored with each device at the factory, and are used for subsequent data encryption and, when necessary, data decryption.

[0094] In this embodiment, the encryption parameter seed c(0) = 0.4, the chaotic range parameter r = 0.3, and the mapping parameter seed t = 0.5 are known, and the number of pixels n of the image to be encrypted is 100.

[0095] Furthermore, chaotic encryption algorithms are used to complete the desensitization process, including:

[0096] (1) Generate the encrypted sequence c(k+1) using the function F(c(k)), as follows:

[0097]

[0098] Where F(c(k)) is the chaotic mapping function, 0≤k≤8n, n is the pixel size of the image to be encrypted, c(k) is the value of the k-th bit in the encryption sequence; r is defined as a value in the range [0,1], which is the known chaotic range parameter. When k=0, c(0) is the known encryption parameter seed.

[0099] In this embodiment, an encrypted sequence [0.4, 0.67, 0.44, 0.82, 0.93, ..., 0.01] consisting of values ​​in the range [0, 1] can be obtained through the chaotic mapping function.

[0100] (2) Obtain the chaotic sequence through the encrypted sequence.

[0101] The length of the chaotic sequence depends on the length of the data that needs to be encrypted.

[0102] The expression for the chaotic sequence d(k) is as follows:

[0103]

[0104] Where G(c(k)) is the mapping function, and the parameter t is a value in the range [0,1], which is the known mapping parameter seed.

[0105] In this embodiment, the encrypted sequence [0.4,0.67,0.44,0.82,0.93,…,0.01] is mapped to the values ​​0 and 1 through G(c(k)) to obtain the chaotic sequence d(k)=[0,1,0,1,1,…,0].

[0106] The encrypted sequence c(k) can be mapped to the values ​​0 and 1 using G(c(k)) to obtain the chaotic sequence d(k).

[0107] (3) Generate a sequence of information to be encrypted based on the image to be encrypted.

[0108] n = i × j, where i and j are the total number of rows and columns of the image to be encrypted, respectively.

[0109] The sequence of information to be encrypted, b(l), corresponds to an eight-bit binary number converted from the grayscale value of the pixel in the p-th row and q-th column of the image to be encrypted, where:

[0110] l=(p-1)×j+q

[0111] Where 0≤l≤n, 0≤p≤i, 0≤q≤j.

[0112] In this embodiment, the sequence of information to be encrypted, b(k), is generated based on the image to be encrypted, b(k) = [1,0,0,1,0,0,0,1,…,1].

[0113] (4) Use chaotic sequences to encrypt the information to be encrypted, and generate an encrypted image based on the encrypted sequence.

[0114] The encrypted location information is the sequence w(l):

[0115]

[0116] in, This is an XOR operation.

[0117] An encrypted image is generated using the sequence w(k), where w(k) corresponds to the eight-bit binary number of the grayscale value of the pixel in the p-th row and q-th column of the encrypted image.

[0118] In this embodiment, the sequence w(k) is obtained as [0,0,1,1,0,1,0,1,…,0]. Based on the first 8 bits of w(k), the pixel grayscale value of the first row and first column in the encrypted image is 53. The pixel grayscale value of the entire encrypted image is obtained by analogy.

[0119] (5) The encrypted image is overwritten with the information at the corresponding position of the original image to obtain the desensitized image, thus achieving the desensitization of key privacy information.

[0120] By overlaying the encrypted portion of the original image with an encrypted image, an image with desensitized key privacy information is obtained.

[0121] (6) The decryption process only requires the same chaotic range parameter r, encryption parameter seed c(0), and mapping parameter seed t. Repeat the above steps (1)-(4) on the image to be decrypted, and the output image is the decrypted image.

[0122] S104. Store the desensitized image into a storage medium.

[0123] Individual users can only view the de-identified images stored on the storage medium. If judicial authorities or relevant departments need to access and view critical privacy information on the storage medium, they only need to go through the decryption steps in S36 to obtain it.

[0124] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the corresponding program can be stored in a computer-readable storage medium.

[0125] It should be noted that although the method operations of the above embodiments are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the order of execution of the described steps may be changed. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0126] Example 2:

[0127] like Figure 4 As shown, this embodiment provides a vehicle recording system with desensitized key privacy information. The system includes a dataset acquisition module 401, a key privacy information location module 402, a key privacy information desensitization module 403, and a desensitized image generation module 404, wherein:

[0128] The dataset acquisition module 401 is used to acquire a dataset of driving recorder images; the driving recorder images include at least one of a human eye and a license plate;

[0129] The key privacy information localization module 402 is used to build a network model architecture, identify human eye and license plate information in the dashcam image through the network model architecture; train the network model architecture with the dataset to obtain a key privacy information localization model, and use the key privacy information localization model to determine the location of key privacy information in the dashcam image to obtain the region to be desensitized; the key privacy information includes face and license plate.

[0130] The key privacy information desensitization module 403 is used to input the area to be desensitized into the privacy information desensitization module based on the chaotic encryption system, generate an encrypted image, and realize the desensitization of the area to be desensitized;

[0131] The desensitized image generation module 404 is used to input the collected images of the actual vehicle driving process into the key privacy information positioning model, input the output area to be desensitized into the privacy information desensitization module, overwrite the information at the corresponding position of the original image with the generated encrypted image to obtain the desensitized image; and store the desensitized image in the storage medium.

[0132] The specific implementation of each module in this embodiment can be found in Embodiment 1 above, and will not be repeated here. It should be noted that the device provided in this embodiment is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above.

[0133] Example 3:

[0134] This embodiment provides a terminal device, which can be a computer, such as... Figure 5 As shown, the system bus 501 connects a processor 502, a memory, an input device 503, a display 504, and a network interface 505. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium 506 and internal memory 507. The non-volatile storage medium 506 stores the operating system, computer programs, and a database. The internal memory 507 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. When the processor 502 executes the computer programs stored in the memory, it implements the vehicle recording method of Embodiment 1 described above, as follows:

[0135] Acquire a dataset of dashcam images; the dashcam images include at least one of a human eye and a license plate;

[0136] A network model architecture is constructed to identify human eye and license plate information in dashcam images. The network model architecture is trained using the dataset to obtain a key privacy information localization model. The key privacy information localization model is used to determine the location of key privacy information in dashcam images, thereby obtaining the region to be desensitized. The key privacy information includes face and license plate.

[0137] The area to be desensitized is input into the privacy information desensitization module based on the chaotic encryption system to generate an encrypted image, thereby achieving the desensitization of the area to be desensitized.

[0138] The images collected during the actual vehicle driving process are input into the key privacy information positioning model. The output area to be desensitized is input into the privacy information desensitization module. The generated encrypted image overwrites the information at the corresponding position of the original image to obtain the desensitized image. The desensitized image is then stored in a storage medium.

[0139] Example 4:

[0140] This embodiment provides a storage medium, which is a computer-readable storage medium, storing a computer program. When the computer program is executed by a processor, it implements the vehicle recording method of Embodiment 1 above, as follows:

[0141] Acquire a dataset of dashcam images; the dashcam images include at least one of a human eye and a license plate;

[0142] A network model architecture is constructed to identify human eye and license plate information in dashcam images. The network model architecture is trained using the dataset to obtain a key privacy information localization model. The key privacy information localization model is used to determine the location of key privacy information in dashcam images, thereby obtaining the region to be desensitized. The key privacy information includes face and license plate.

[0143] The area to be desensitized is input into the privacy information desensitization module based on the chaotic encryption system to generate an encrypted image, thereby achieving the desensitization of the area to be desensitized.

[0144] The images collected during the actual vehicle driving process are input into the key privacy information positioning model. The output area to be desensitized is input into the privacy information desensitization module. The generated encrypted image overwrites the information at the corresponding position of the original image to obtain the desensitized image. The desensitized image is then stored in a storage medium.

[0145] It should be noted that the computer-readable storage medium in this embodiment can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0146] In summary, the method provided by this invention mainly includes: acquiring a dataset of human eye and license plate images; dividing the dataset into a training set and a test set according to a certain ratio based on existing calibrated images containing human eye and license plate information; training a network model architecture to obtain a key privacy information localization model; training the parameters of the network model architecture, and after the parameters converge, obtaining a model that can output the coordinate information of the predicted bounding boxes of key privacy information; inputting the collected images of the vehicle during normal driving into the key privacy information localization model, outputting the predicted bounding boxes of the face and license plate to be desensitized; inputting the coordinates of the predicted bounding boxes into the key privacy information desensitization module for desensitization processing, generating desensitized images and storing them in a data storage medium, thereby achieving the desensitization of key privacy information.

[0147] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A method for desensitizing key privacy information in a vehicle recording system, characterized in that, The method includes: Obtain a dataset of dashcam images; the dashcam images include at least one of a human eye and a license plate; A network model architecture is constructed, comprising a preprocessing module, a feature extraction skeleton module, a feature fusion pyramid module, and output Loss and NMS modules. The preprocessing module is used to preprocess the dashcam images. The feature extraction skeleton module uses a Focus structure and the CSP structure from CSPNet to extract features from the preprocessed image. The feature fusion pyramid module is a fusion structure of FPN and PAN structures, used to integrate parameters and extract features from feature layers of different sizes from different backbone layers, thereby achieving [the desired feature extraction]. Further information is extracted from the extracted features; the loss function uses GIOU_Loss as the loss function for the target detection box; the NMS module is used to filter the predicted target detection boxes; a key privacy information localization model is obtained by training the network model architecture using the dataset, and the key privacy information localization model is used to locate the key privacy information in the dashcam image; the key privacy information includes the face and license plate; based on the location of the human eye, the face is located and labeled according to the ratio of the human eye and the face to determine the location of the face; the location information of the face and the license plate is combined to obtain the region to be desensitized; The region to be desensitized is input into the privacy information desensitization module based on the chaotic encryption system. An encryption sequence is generated according to the chaotic mapping function and the pixel size of the region to be desensitized. A chaotic sequence is obtained by mapping the encryption sequence. An information sequence to be encrypted is generated according to the region to be desensitized. Encryption is achieved by performing an XOR operation between the chaotic sequence and the information to be encrypted. An encrypted image is generated according to the encrypted sequence to achieve desensitization of the region to be desensitized. The images collected during the actual vehicle driving process are input into the key privacy information positioning model. The output area to be desensitized is input into the privacy information desensitization module. The generated encrypted image overwrites the information at the corresponding position of the original image to obtain the desensitized image. The desensitized image is then stored in a storage medium. The loss function GIOU_Loss is: ； ； In the formula, intersection A is the area of ​​the overlapping part of the real object detection box and the predicted object detection box, and union B is the total area occupied by the real object detection box and the predicted object detection box.

2. The driving recording method according to claim 1, characterized in that, The Focus structure is used to adjust the feature map size; The CSP structure includes convolutional layers and residual units, which are used to extract features from the input feature map. The FPN structure is an upsampling process that conveys strong semantic information from top to bottom. The PAN structure is a bottom-up feature pyramid that conveys strong positioning information from the bottom up.

3. The driving recording method according to claim 1, characterized in that, The preprocessing includes Mosaic data augmentation, cmBN, and adaptive image scaling, wherein the Mosaic data augmentation includes random arrangement, random scaling, and random cropping of the image.

4. A vehicle recording system for desensitizing key privacy information, characterized in that, The system includes: A dataset acquisition module is used to acquire a dataset of driving recorder images; the driving recorder images include at least one of a human eye and a license plate; A key privacy information location module is used to build the network model architecture. This network model architecture includes a preprocessing module, a feature extraction skeleton module, a feature fusion pyramid module, and output Loss and NMS modules. The preprocessing module is used to preprocess the dashcam images. The feature extraction skeleton module uses a Focus structure and the CSP structure from CSPNet to extract features from the preprocessed image. The feature fusion pyramid module uses a fusion structure of FPN and PAN structures to integrate parameters and features from different backbone layers of different sizes. The extraction process is as follows: More information is extracted based on the extracted features; the loss function uses GIOU_Loss as the loss function for the target detection box; the NMS module is used to filter the predicted target detection boxes; a key privacy information localization model is obtained by training the network model architecture using the dataset; the key privacy information localization model is used to locate key privacy information in the dashcam image; the key privacy information includes the face and license plate; based on the location of the human eye, the face is located and labeled according to the ratio of the human eye and the face to determine the location of the face; the location information of the face and the license plate is combined to obtain the region to be desensitized. The critical privacy information desensitization module is used to input the area to be desensitized into the privacy information desensitization module based on the chaotic encryption system, generate an encryption sequence according to the chaotic mapping function and the pixel size of the area to be desensitized, obtain a chaotic sequence by mapping the encryption sequence, generate an information sequence to be encrypted according to the area to be desensitized, achieve encryption by performing an XOR operation between the chaotic sequence and the information to be encrypted, and generate an encrypted image according to the encrypted sequence to achieve desensitization of the area to be desensitized. The desensitized image generation module is used to input the collected images of the actual vehicle driving process into the key privacy information positioning model, input the output area to be desensitized into the privacy information desensitization module, overwrite the information at the corresponding position of the original image with the generated encrypted image to obtain the desensitized image; and store the desensitized image in the storage medium. The loss function GIOU_Loss is: ； ； In the formula, intersection A is the area of ​​the overlapping part of the real object detection box and the predicted object detection box, and union B is the total area occupied by the real object detection box and the predicted object detection box.

5. A terminal device, comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the driving recording method according to any one of claims 1-3.

6. A storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the vehicle recording method according to any one of claims 1-3.

Images