Vehicle end image desensitization method and system, storage medium and vehicle-mounted terminal

Abstract

The invention provides a vehicle end image desensitization method and system, a storage medium and a vehicle-mounted terminal. The vehicle end image desensitization method comprises the following steps: acquiring a vehicle end image in real time, and transmitting the vehicle end image to a desensitization model; performing convolution operation on the vehicle end image through a desensitization model to obtain semantic features and edge features, and obtaining target desensitization features based on the semantic features and the edge features; further obtaining desensitized target data at least comprising a target category and a target position based on target desensitization feature regression; and performing fuzzy processing on the target position, and marking a corresponding target category at the target position after the fuzzy processing so as to obtain a target desensitized video and output the target desensitized video to the terminal. According to the invention, desensitization processing can be carried out while the vehicle end image is acquired in real time, and the target needing desensitization can be accurately identified and positioned, so that the desensitization accuracy is improved, and the data security is guaranteed.

Description

Technical Field

[0001] The present application relates to the technical field of data desensitization, and particularly relates to a vehicle-end image desensitization method, system, storage medium, and vehicle-mounted terminal. Background Art

[0002] An intelligent connected vehicle refers to the organic combination of the vehicle network and intelligent vehicles. It is equipped with advanced vehicle-mounted sensors, controllers, actuators, and other devices, and integrates modern communication and network technologies to achieve intelligent information exchange and sharing between vehicles and people, roads, and the background, realize safe, comfortable, energy-saving, and efficient driving, and ultimately can replace humans to operate a new generation of vehicles.

[0003] Under the trend of intelligentization and electrification, the form of automotive products has changed from traditional travel tools to mobile intelligent terminals. In addition to vehicle data itself, intelligent connected vehicles also include urban traffic data and user personal data. Vehicles will become the center for generating and connecting a large amount of data. Inside the vehicle, a large amount of biometric data with accurate personal characteristics is integrated through devices such as in-vehicle cameras, microphones, and cockpit sensors. Outside the vehicle, information about surrounding pedestrians can be continuously obtained through vehicle-mounted sensors. Once these data are leaked, tampered with, or illegally shared, risks such as eavesdropping and identity theft are extremely likely to occur.

[0004] Existing desensitization technologies have low detection accuracy in complex scenarios, resulting in the inability to accurately desensitize vehicle-end images. Summary of the Invention

[0005] To solve the above technical problems, the present application provides a vehicle-end image desensitization method, system, storage medium, and vehicle-mounted terminal with relatively high detection accuracy.

[0006] Specifically, the present application provides a vehicle-end image desensitization method, including the following steps:

[0007] Real-time obtain a vehicle-end image and transfer the vehicle-end image to a desensitization model.

[0008] Perform a convolution operation on the vehicle-end image through the desensitization model to obtain semantic features and edge features, and obtain target desensitization features based on the semantic features and edge features.

[0009] Regress based on the target desensitization features to obtain desensitization target data, where the desensitization target data at least includes a target category and a target position.

[0010] And, perform blurring processing on the target position, and mark the corresponding target category at the blurred target position to obtain a target desensitization video and output it to the terminal.

[0011] In the above technical solution, it is possible to perform desensitization processing while obtaining the vehicle-end image in real time, effectively preventing the leakage of personal biometric data and pedestrian information. Moreover, by combining multi-layer feature extraction and regression algorithms, the target to be desensitized can be accurately identified and located, improving the accuracy of desensitization and further ensuring data security.

[0012] Further, the obtaining of the vehicle-end image includes:

[0013] Real-time capture of vehicle-end video frames and decoding the vehicle-end video frames into vehicle-end images.

[0014] In the above technical solution, by real-time capturing and decoding video frames, the real-time update of image data can be ensured.

[0015] Among them, the video frames can select different resolutions and frame rates according to needs to adapt to different application scenarios; for example, a higher frame rate can be selected when driving at high speed to capture more details, while a lower frame rate can be selected when driving at low speed to reduce the computational burden.

[0016] In addition, the video decoding process can be compatible with multiple video compression formats, ensuring the ability to process video streams from different devices and systems, and improving the versatility and compatibility of the system.

[0017] Further, the obtaining of the semantic features and edge features includes:

[0018] Normalize the vehicle-end image to a preset size to obtain multiple sample groups based on the normalized vehicle-end image.

[0019] Extract basic features from the multiple sample groups based on residual blocks and multiple convolutional layers.

[0020] And extract semantic features and edge features from the basic features based on multiple convolutional layers.

[0021] In the above technical solution, through the combination of residual blocks and multiple convolutional layers, the features of the image can be deeply and widely extracted. The skip connection of the residual block ensures the depth of the network, while the multi-level processing of the convolutional layer ensures the breadth of feature extraction, thus improving the quality and richness of the features; the normalized standard size enables the network to process images of different sizes more consistently, and the segmentation of the sample groups enables the network to focus on the local details of the image, improving the accuracy of feature extraction; through multi-level convolutional processing, the network can extract various types of features, including low-level features (such as edges) and high-level features (such as semantic information), and these diverse features provide a rich information basis for subsequent desensitization processing.

[0022] Further, the obtaining of the target desensitization features includes:

[0023] Connect the semantic features and the edge features through upsampling, and combine the edge features with the upsampled semantic features through lateral connection to output the target desensitized features.

[0024] In the above technical solution, through upsampling and lateral connection, multi-scale features can be effectively fused. The semantic features provide high-level category information, while the edge features provide low-level shape and contour information. After the two are combined, the target desensitized features can contain more comprehensive information, improving the accuracy and effect of the desensitization process.

[0025] Further, the obtaining of the desensitized target data includes:

[0026] Generate corresponding anchor boxes based on the target desensitized features, and obtain the class probability distribution and bounding box regression parameters corresponding to each anchor box.

[0027] Obtain the target class corresponding to the anchor box based on the class probability distribution, and at the same time apply the bounding box regression parameters to the anchor box to obtain the bounding box coordinates as the target position.

[0028] In the above technical solution, through anchor boxes and bounding box regression, high-precision object detection can be achieved. The anchor boxes cover the possible target areas, and the bounding box regression parameters further refine these areas to ensure the accuracy of the target position and boundary; the class probability distribution allows the system to identify multiple target classes, not limited to a single class, which is very important for various application scenarios in intelligent connected vehicles (such as face recognition, license plate recognition, etc.) and can adapt to different types of object detection requirements.

[0029] Further, the obtaining of the target desensitized video includes:

[0030] Obtain the confidence of each anchor box, and remove the overlapping anchor boxes based on the confidence to obtain the final anchor boxes.

[0031] In the above technical solution, by removing the overlapping anchor boxes, redundant detection results can be significantly reduced, improving the cleanliness and simplicity of the detection.

[0032] Further, the obtaining of the target desensitized video further includes:

[0033] Perform blurring processing on the target position corresponding to the final anchor box using a preset blurring algorithm, and add the class label corresponding to the target class to the blurred target position to obtain a desensitized image.

[0034] Merge the desensitized image and the vehicle-end image to form desensitized fusion data, and encode the desensitized fusion data to obtain the target desensitized video.

[0035] In the above technical solution, the details of the target can be effectively hidden through blurring processing to protect privacy. The selection of the blurring algorithm and the parameter settings can be adjusted according to specific requirements to ensure the best privacy protection effect. Adding category labels to the target positions after blurring processing can retain the category information of the target while protecting privacy, facilitating subsequent analysis and processing, and ensuring that the desensitization processing does not affect the recognition and classification of the target.

[0036] Furthermore, based on the same concept, the present application also provides a vehicle-end image desensitization system, including:

[0037] An acquisition module, configured to acquire vehicle-end images in real time and transfer the vehicle-end images to the desensitization model.

[0038] A desensitization module, configured to obtain target desensitization features based on the vehicle-end images through the desensitization model, and regress to obtain the target category and target position based on the target desensitization features, so as to perform blurring processing on the target position, and at the same time label the corresponding target category at the blurred target position to obtain a target desensitized video.

[0039] And an output module, configured to output the target desensitized video to the terminal.

[0040] In the above technical solution, the system can acquire vehicle-end images in real time and perform rapid processing through the desensitization model to generate a target desensitized video, which ensures the real-time performance and stability of the system in a high-speed moving vehicle-mounted environment and meets the requirements of real-time monitoring and privacy protection.

[0041] Furthermore, based on the same concept, the present application also provides a storage medium, in which a computer program is stored, and the computer program is set to execute the vehicle-end image desensitization method when running.

[0042] Furthermore, based on the same concept, the present application also provides a vehicle-mounted terminal, the vehicle-mounted terminal includes a processor and a memory, and at least one instruction, at least one program, a code set or an instruction set is stored in the memory, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the vehicle-end image desensitization method.

[0043] Compared with the prior art, the beneficial effects of the present application are as follows:

[0044] This application first obtains vehicle - end images in real - time and transfers the vehicle - end images to a desensitization model. Then, the desensitization model performs convolution operations on the vehicle - end images to obtain semantic features and edge features, and based on the semantic features and edge features, obtains target desensitization features. Further, based on the target desensitization features, regression is performed to obtain desensitized target data including at least a target category and a target location. Then, the target location is blurred, and the corresponding target category is marked at the blurred target location to obtain a target desensitized video and output it to the terminal. This application can perform desensitization processing while obtaining vehicle - end images in real - time, and can accurately identify and locate the targets that need to be desensitized, improving the accuracy of desensitization and ensuring data security. Description of the Drawings

[0045] Figure 1 It is a flowchart of the vehicle - end image desensitization method described in this application.

[0046] Figure 2 It is a data - flow diagram of the vehicle - end image desensitization method described in this application.

[0047] Figure 3 It is a framework diagram of the vehicle - end image desensitization system described in this application. Detailed Embodiments

[0048] The following further describes in detail a vehicle - end image desensitization method, system, storage medium, and vehicle - mounted terminal of this application in combination with specific embodiments and the accompanying drawings.

[0049] This application provides a vehicle - end image desensitization method. Please refer to Figure 1 and Figure 2 , which includes the following steps S100 - S300.

[0050] In some embodiments, the APP that collects vehicle - end images transfers the vehicle - end images to the desensitization algorithm library (i.e., the desensitization model) through, for example, the JNI layer. Among them, after receiving the vehicle - end images, JNI calls the DDT algorithm library for algorithm model inference, including at least feature extraction, and then outputs the target category and target location through the head layer. Then, the target category and target location are transferred to the APP side through the JNI layer, so that the APP blurs the corresponding data according to the target location and outputs the desensitized data to the mobile phone or other mobile terminals (i.e., the terminal).

[0051] Among them, the JNI layer is an interface layer and a bridge for data transmission. Those skilled in the art can choose a suitable way to transfer the vehicle - end images and other data according to actual application requirements.

[0052] The following details steps S100 - S300.

[0053] Among them, step S100: Obtain the vehicle-end image in real time and transfer the vehicle-end image into the desensitization model.

[0054] Furthermore, the obtaining of the vehicle-end image includes:

[0055] Capture the vehicle-end video frame in real time and decode the vehicle-end video frame into a vehicle-end image.

[0056] In some embodiments, obtain the camera data stream (i.e., the vehicle-end video frame) from the vehicle-mounted camera, and then call, such as MediaCodec, to decode the vehicle-end video frame to obtain a vehicle-end image in a preset format (such as jpeg).

[0057] The MediaCodec is an API provided by the Android platform for audio and video encoding and decoding, which allows developers to directly interact with the hardware codec to achieve efficient audio and video encoding and decoding operations.

[0058] Specifically: Create a MediaCodec instance, specify the decoder type, configure the input and output formats of MediaCodec, input the vehicle-end video frame into MediaCodec for decoding, obtain the decoded image data from MediaCodec. At this time, the image data is usually in YUV format. Convert the YUV-format image data into RGB data, and then use an RGB encoder to encode the RGB data into a vehicle-end image in jpeg format.

[0059] Among them, in other embodiments, other decoding methods can also be adopted, and it can also be decoded into other formats, which is not limited thereto.

[0060] Among them, it should be noted that, in some embodiments, an 8255 embedded computing platform is adopted. The Hypervisor dual system (QNX + Android) on this computing platform can meet the requirements of different application scenarios; however, the sharing of camera data between the QNX and Android systems has not been realized yet, and the function of splitting the video stream into two by CamereHAL (Camera Hardware Abstraction Layer) has not been realized either. Therefore, the desensitization system in this application is deployed on the Android side.

[0061] In the above technical solution, by capturing and decoding the video frame in real time, the real-time update of the image data can be ensured.

[0062] Among them, the video frame can select different resolutions and frame rates according to needs to adapt to different application scenarios; for example, a higher frame rate can be selected when driving at high speed to capture more details, while a lower frame rate can be selected when driving at low speed to reduce the computational burden.

[0063] In addition, the video decoding process can be compatible with a variety of video compression formats to ensure that it can process video streams from different devices and systems, improving the versatility and compatibility of the system.

[0064] After the vehicle-end image is input into the desensitization model, step S200 can be executed.

[0065] Among them, step S200: perform convolution operations on the vehicle-end image through the desensitization model to obtain semantic features and edge features, and obtain target desensitization features based on the semantic features and edge features.

[0066] Further, the obtaining of semantic features and edge features includes:

[0067] Normalize the vehicle-end image to a preset size to obtain multiple sample groups based on the normalized vehicle-end image.

[0068] Extract basic features from the multiple sample groups based on residual blocks and multiple convolutional layers.

[0069] And extract semantic features and edge features from the basic features based on multiple convolutional layers.

[0070] In some embodiments, the vehicle-end image is normalized to 8*3*800*640 (n*c*h*w), where n is the number of pictures in a sample group, c is the three RGB channels of a single picture, and h*w is the pixel size of the picture.

[0071] It should be noted that those skilled in the art can also select other sizes for normalization according to actual application requirements, and are not limited to this.

[0072] Furthermore, by introducing residual blocks, the problem of gradient disappearance in deep networks is solved, enabling the network to be deeper and extract richer features; multiple small convolutional kernels (such as 3*3) can be stacked to increase the network depth, extract multi-level features, and extract features in parallel through convolutional kernels of different sizes to increase the network width and expressive ability.

[0073] Preferably, use the Backbone (such as ResNet) in the YOLO object detection model to extract multi-level basic features from the vehicle-end image.

[0074] Further, fuse the image features through the FPN (Feature Pyramid Network) structure; among them, use multiple convolutional layers to extract deep semantic features. These convolutional layers usually have more channels and contain deeper network structures. Further reduce the size of the feature map through the pooling layer and increase the depth of the feature map.

[0075] Further, use shallower convolutional layers to extract shallow edge features. These convolutional layers usually have fewer channels, and the size of the feature map is larger. Do not perform downsampling and keep the size of the feature map unchanged to retain more spatial information.

[0076] In the above technical solution, through the combination of residual blocks and multiple convolutional layers, the features of the image can be extracted deeply and extensively. The skip connection of the residual block ensures the depth of the network, while the multi-level processing of the convolutional layer ensures the breadth of feature extraction, thereby improving the quality and richness of the features; the normalized standard size enables the network to process images of different sizes more consistently, and the segmentation of the sample group enables the network to focus on the local details of the image and improves the accuracy of feature extraction; through multi-level convolutional processing, the network can extract various types of features, including low-level features (such as edges) and high-level features (such as semantic information). These diverse features provide a rich information basis for subsequent desensitization processing.

[0077] Further, the obtaining of the target desensitized features includes:

[0078] Connect the semantic features and the edge features through upsampling, and combine the edge features with the upsampled semantic features through lateral connection to output and obtain the target desensitized features.

[0079] In some embodiments, starting from the deep semantic features, connect with the shallow edge features through upsampling, gradually restore the spatial details, and combine the shallow edge features with the upsampled deep semantic features through lateral connection to obtain a multi-scale feature map, that is, the target desensitized features.

[0080] In the above technical solution, through upsampling and lateral connection, multi-scale features can be effectively fused. The semantic features provide high-level category information, while the edge features provide low-level shape and contour information. After the two are combined, the target desensitized features can contain more comprehensive information, improving the accuracy and effect of desensitization processing.

[0081] After obtaining the target desensitized features, step S300 can be executed.

[0082] Among them, step S300: Obtain desensitized target data based on the target desensitization feature regression, where the desensitized target data includes at least a target category and a target location.

[0083] Further, the obtaining of the desensitized target data includes:

[0084] Generate corresponding anchor boxes based on the target desensitization feature, and obtain the class probability distribution and bounding box regression parameters corresponding to each anchor box.

[0085] Obtain the target category corresponding to the anchor box based on the class probability distribution, and at the same time apply the bounding box regression parameters to the anchor box to obtain the bounding box coordinates as the target location.

[0086] In some embodiments, a set of predefined anchor boxes, such as rectangular boxes, are generated at each position of each target desensitization feature; each anchor box has different sizes and aspect ratios (i.e., the bounding box regression parameters) to cover targets of different sizes and shapes.

[0087] Further, for each anchor box, predict its corresponding class probability and bounding box offset through the head layer.

[0088] Among them, for each anchor box, a vector is output, representing the probability of belonging to each class, and at the same time four values are output, representing the coordinate offsets of the bounding box (i.e., the target location).

[0089] In the above technical solution, through anchor boxes and bounding box regression, high-precision object detection can be achieved. The anchor boxes cover the possible target areas, and the bounding box regression parameters further refine these areas to ensure the accuracy of the target location and boundaries; the class probability distribution allows the system to identify multiple target categories, not limited to a single category, which is very important for various application scenarios in intelligent connected vehicles (such as face recognition, license plate recognition, etc.) and can adapt to different types of object detection requirements.

[0090] After obtaining the desensitized target data, step S400 can be executed.

[0091] Among them, step S400: Blur the target location, and mark the corresponding target category at the blurred target location to obtain a target desensitized video and output it to the terminal.

[0092] Further, the obtaining of the target desensitized video includes:

[0093] Obtain the confidence of each anchor box, and remove overlapping anchor boxes based on the confidence to obtain the final anchor boxes.

[0094] Further, by setting a confidence threshold and an IOU (Intersection over Union) threshold, detecting boxes with low confidence and high overlap are filtered out, and the predicted offset is applied to the anchor box to obtain the final bounding box coordinates (i.e., the target position, in xywh format).

[0095] In the above technical solution, by removing overlapping anchor boxes, redundant detection results can be significantly reduced, improving the cleanliness and simplicity of detection.

[0096] Further, obtaining the target desensitized video further includes:[[]]END]]

[0097] Using a preset blurring algorithm to blur the target position corresponding to the final anchor box, and adding a class label corresponding to the target category to the blurred target position to obtain a desensitized image.

[0098] Merging the desensitized image and the vehicle-end image to form desensitized fusion data, and encoding the desensitized fusion data to obtain a target desensitized video.

[0099] In some embodiments, a blurring algorithm (such as Gaussian blur, mean blur, or median blur) is used to process the area where the target position is located, and the degree of blurring can be adjusted according to specific requirements; on the blurred target position, a text label can be added to indicate the category of the target; wherein, those skilled in the art pre-select appropriate font, size, color, and position to display the category label, ensuring that the label is clearly visible and does not affect the overall visual perception of the image.

[0100] Further, merging the blurred and labeled image area with the remaining part of the original vehicle-end image to form a final output image, i.e., the desensitized fusion data.

[0101] Further, calling, such as MediaCodec, to encode the desensitized fusion data to obtain a target desensitized video, and uploading the target desensitized video to a terminal (such as a mobile phone or other mobile device) for display.

[0102] In the above technical solution, through blurring processing, the detailed information of the target can be effectively hidden to protect privacy. The selection and parameter settings of the blurring algorithm can be adjusted according to specific requirements to ensure the best privacy protection effect; adding a category label to the blurred target position can, while protecting privacy, retain the category information of the target, facilitating subsequent analysis and processing, and ensuring that the desensitization process does not affect the recognition and classification of the target.

[0103] Further, based on the same concept, please refer to Figure 3 , this application also provides a vehicle-end image desensitization system, including:[[]]END]]

[0104] An acquisition module, configured to acquire vehicle-end images in real time and transmit the vehicle-end images to a desensitization model.

[0105] A desensitization module, configured to obtain target desensitization features based on the vehicle-end images through the desensitization model, and perform regression based on the target desensitization features to obtain a target category and a target location, so as to perform blurring processing on the target location, and at the same time label the corresponding target category at the blurred target location to obtain a target desensitized video.

[0106] And an output module, configured to output the target desensitized video to a terminal.

[0107] It should be noted that the vehicle-end image desensitization system adopts the vehicle-end image desensitization method described in the above embodiments, and the specific implementation manner of this system is the same as that of the method, which will not be elaborated here.

[0108] In the above technical solution, the system can acquire vehicle-end images in real time and perform fast processing through a desensitization model to generate a target desensitized video, which ensures the real-time performance and stability of the system in a high-speed moving vehicle-mounted environment and meets the requirements of real-time monitoring and privacy protection.

[0109] Furthermore, based on the same concept, the present application also provides a storage medium, in which a computer program is stored, and the computer program is configured to execute the vehicle-end image desensitization method when running.

[0110] In some embodiments, the storage medium stores a number of computer programs for enabling a vehicle-mounted terminal to execute all or part of the steps of the methods described in various embodiments of the present application.

[0111] The medium may include various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk, or an optical disc.

[0112] Furthermore, based on the same concept, the present application also provides a vehicle-mounted terminal, which includes a processor and a memory. At least one instruction, at least one program, a code set, or an instruction set is stored in the memory, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the vehicle-end image desensitization method.

[0113] In some embodiments, the memory and the processor are interconnected through a bus; the processor may be one or more CPUs. When the processor is a single CPU, the CPU may be a single-core CPU or a multi-core CPU, and the processor is used to control each functional module of the vehicle-mounted terminal and process signals.

[0114] The memory includes, but is not limited to, RAM (Random Access Memory), ROM (Read-Only Memory), EPROM (Erasable Programmable Read-Only Memory), CD-ROM (Compact Disc Read-Only Memory). The memory is used to store computer programs, operating systems, various applications and data, such as storing the computer program for implementing the vehicle-end image desensitization method.

[0115] In summary, the present application provides a vehicle-end image desensitization method, system, storage medium and vehicle-mounted terminal. First, a vehicle-end image is obtained in real time and transmitted to a desensitization model. Then, the desensitization model performs a convolution operation on the vehicle-end image to obtain semantic features and edge features, and based on the semantic features and edge features, target desensitization features are obtained. Further, based on the target desensitization features, regression is performed to obtain desensitization target data including at least a target category and a target position. Then, the target position is blurred, and the corresponding target category is marked at the blurred target position to obtain a target desensitized video for output to the terminal. The present application can perform desensitization processing while obtaining the vehicle-end image in real time, and can accurately identify and locate the targets that need to be desensitized, improving the accuracy of desensitization and ensuring data security.

[0116] Although example embodiments have been described herein with reference to the drawings, it should be understood that the above example embodiments are merely exemplary and are not intended to limit the scope of the present application. Those of ordinary skill in the art can make various changes and modifications therein without departing from the scope and spirit of the present application. All such changes and modifications are intended to be included within the scope of the present application as claimed in the appended claims.

[0117] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0118] In several embodiments provided by this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed.

[0119] Each component embodiment of this application can be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. Those skilled in the art should understand that a microprocessor or a digital signal processor (DSP) can be used in practice to implement some or all of the functions of some modules according to the embodiments of this application. This application can also be implemented as a device program (such as a computer program and a computer program product) for executing part or all of the methods described herein. Such a program implementing this application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, or provided on a carrier signal, or in any other form.

[0120] It should be noted that in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the existence of additional identical elements in the process, method, article or device comprising the element.

[0121] Although the description of this application is made in conjunction with the above specific embodiments, it is obvious that those familiar with the technical field can make many substitutions, modifications and changes based on the above content. Therefore, all such substitutions, improvements and changes are included within the spirit and scope of the appended claims.

Claims

1. A vehicle-side image desensitization method, characterized in that: The following steps are involved: Acquire vehicle-side images in real time, and transfer the vehicle-side images to the desensitization model; Performing a convolution operation on the vehicle-side image through the desensitization model to obtain semantic features and edge features, and obtaining target desensitization features based on the semantic features and edge features; Obtaining desensitized target data based on the target desensitization feature regression, wherein the desensitized target data at least includes a target category and a target position; And, the target position is blurred, and the corresponding target category is marked at the blurred target position to obtain a target desensitized video output to the terminal.

2. The vehicle-side image desensitization method according to claim 1, characterized in that: The obtaining of the vehicle-side image comprises: The vehicle-side video frames are captured in real time, and the vehicle-side video frames are decoded into vehicle-side images.

3. The vehicle-side image desensitization method according to claim 2, characterized in that: The obtaining of semantic features and edge features comprises: Normalizing the vehicle-side image to a preset size to obtain multiple sample groups based on the normalized vehicle-side image; Extracting basic features from the multiple sample groups based on a residual block and multiple convolutional layers; And, semantic features and edge features are extracted from the basic features based on multiple convolutional layers.

4. The vehicle-side image desensitization method according to claim 3 is characterized in that: The obtaining of the target desensitized feature includes: connecting the semantic feature with the edge feature through upsampling, and combining the edge feature with the upsampled semantic feature through lateral connection to output the obtained target desensitized feature.

5. The vehicle-side image desensitization method according to claim 4, characterized in that: The obtaining of desensitized target data includes: generating corresponding anchor frames based on the target desensitized features, and obtaining category probability distribution and bounding box regression parameters corresponding to each anchor frame; The target category corresponding to the anchor box is obtained based on the category probability distribution, and the bounding box regression parameters are applied to the anchor box to obtain bounding box coordinates as the target position.

6. The vehicle-side image desensitization method according to claim 5, characterized in that: The obtaining of the target desensitized video includes: obtaining the confidence of each anchor frame, removing overlapping anchor frames based on the confidence, and obtaining a final anchor frame.

7. The vehicle-side image desensitization method according to claim 6, characterized in that: The step of obtaining the target desensitized video further includes: using a preset fuzzy algorithm to perform a fuzzy process on the target position corresponding to the final anchor frame, and adding a category label corresponding to the target category to the fuzzy target position to obtain a desensitized image; The desensitized image and the vehicle-side image are merged to form desensitized fusion data, and the desensitized fusion data is encoded to obtain a target desensitized video.

8. A system using the vehicle-side image desensitization method according to any one of claims 1 to 7, characterized in that: include: An acquisition module, used for acquiring vehicle-side images in real time and transferring the vehicle-side images to a desensitization model; A desensitization module, used to obtain a target desensitization feature based on the vehicle-side image through the desensitization model, and regress the target category and target position based on the target desensitization feature to blur the target position, and mark the corresponding target category at the blurred target position to obtain a target desensitization video; And an output module, used to output the target desensitized video to a terminal.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the vehicle-side image desensitization method as described in any one of claims 1 to 7 when running.

10. A vehicle-mounted terminal, characterized in that: The vehicle-mounted terminal includes a processor and a memory, and the memory stores at least one instruction, at least one program, a code set or an instruction set. The at least one instruction, the at least one program, the code set or the instruction set are loaded and executed by the processor to implement the vehicle-end image desensitization method as described in any one of claims 1-7.

Images