A vehicle end perception sensitive target desensitization method and device and storage medium

Abstract

This application discloses a method, apparatus, and storage medium for desensitizing sensitive targets perceived by a vehicle, relating to the field of image recognition technology. The method includes: inputting an acquired vehicle image into a pre-constructed information perception model for sensitive target perception; determining a first detection box based on a target detection network, a second detection box based on a target tracking module, a third detection box based on an optical flow motion estimation module, and a fourth detection box based on a state prediction module; determining weights corresponding to each detection box based on a pre-set weight adjustment strategy, and performing weighted fusion on the first, second, third, and fourth detection boxes to output a final detection box; and locating sensitive target regions in the vehicle image based on the final detection boxes, and desensitizing the sensitive target regions based on a preset desensitization method. Therefore, this application can effectively suppress detection box jitter, reduce the target false negative rate, and improve positioning stability.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, and in particular to a method, apparatus, and storage medium for desensitizing sensitive targets perceived by vehicles. Background Technology

[0002] With the rapid development of the intelligent vehicle industry, in-vehicle image acquisition devices such as front-view cameras, surround-view cameras, and dashcams continuously collect image data containing sensitive targets such as faces and license plates during vehicle operation. Intelligent vehicles involve a large amount of vehicle operation information and user privacy during massive data interactions with the cloud, mobile terminals, and in-vehicle devices. A lack of effective regulation could lead to significant security risks.

[0003] Currently, vehicle-side sensitive target perception mainly relies on deep learning object detection models, with single-stage detection algorithms, such as the YOLO series, demonstrating good detection performance in general scenarios. However, in practical vehicle applications, issues arise such as insufficient vehicle-side computing power and weak robustness in complex environments. Furthermore, general object detection models typically have a large number of parameters and high computational complexity, and lightweight research specifically optimized for vehicle-side sensitive targets such as faces and license plates is still insufficient. In addition, most existing detection methods adopt single-frame object detection, relying mainly on the object detection network to directly output detection results, lacking effective utilization of video temporal information. In real-world vehicle scenarios, due to complex factors such as high-speed vehicle movement, short-term target occlusion, low brightness at night, and motion blur, relying solely on detection models easily leads to problems such as detection box jitter, missed target detection, and unstable localization.

[0004] For example, publication number CN117894008A, invention titled "An Efficient License Plate Detection Method Based on Gold-YOLO and PARSeq Cascade," includes the following steps: Step 1, real-time acquisition of vehicle images on traffic segments, followed by encoding and remote transmission of the vehicle images; Step 2, decoding of the encoded vehicle images and subsequent data augmentation processing; Step 3, processing of the vehicle images using a trained GoldYOLO network model to output the location of the license plate region; Step 4, image augmentation processing of the license plate region to highlight text information; correction of tilted license plate regions to obtain the license plate image; Step 5, processing of the license plate image using a trained PARSeq network model to output the license plate number.

[0005] For example, the invention disclosed in CN119625688A, entitled "Efficient Real-Time Intelligent Vehicle Detection and Recognition Method and System Based on YOLO-BOS," describes a method that trains a YOLOBOS model using vehicle sample images to obtain a vehicle detection and recognition model. The YOLOBOS model includes: an input terminal for preprocessing the input image; a backbone network for extracting features from the preprocessed image; a head network for fusing the extracted image features; and a prediction output terminal for classifying and predicting the fused image features. The backbone network utilizes a two-layer routing mechanism in the BRSA attention mechanism to adjust the weights of each spatial location in the image to filter and enhance key features. The head network uses full-dimensional dynamic convolution to expand the model's receptive field and perception range to capture relevant features during feature fusion. The vehicle detection and recognition model is then used to extract the category and bounding box information of the target vehicle in the image to be detected.

[0006] Existing vehicle-side desensitization technologies generally focus on the desensitization process itself, employing a single method to handle sensitive targets. This not only lacks flexible configuration capabilities for different application scenarios but also fails to provide effective quantitative evaluation of the desensitization effect. In practical applications, simply completing the desensitization operation without assessing whether the desensitization is sufficient, whether there are omissions or over-processing, makes it difficult to meet the requirement of verifiability of the desensitization results.

[0007] There is currently no effective solution to the technical problems of insufficient robustness of target detection methods in the complex environment of vehicles and unstable detection and positioning in the existing technologies. Summary of the Invention

[0008] The embodiments of this disclosure provide a method, apparatus, and storage medium for desensitizing sensitive targets perceived by a vehicle, in order to at least solve the technical problems of insufficient robustness of target detection methods in the perception of complex environments on the vehicle side and unstable detection and positioning in the prior art.

[0009] According to one aspect of the present disclosure, a method for desensitizing sensitive targets perceived by a vehicle is provided, comprising: inputting an acquired vehicle-side image into a pre-constructed information perception model for sensitive target perception, wherein the information perception model includes a target detection network for detecting sensitive targets, a target tracking module for recording the trajectory of sensitive targets, an optical flow motion estimation module for estimating the cross-frame motion trend of sensitive targets, and a state prediction module for predicting the future motion state of sensitive targets, and the vehicle-side image includes multiple image frames; determining a first detection box corresponding to a sensitive target within each image frame based on the target detection network, determining a second detection box corresponding to a sensitive target within each image frame based on the target tracking module, and determining a second detection box corresponding to a sensitive target within each image frame based on the optical flow motion estimation module. The system includes a third detection box corresponding to the sensitive target within each image frame and a fourth detection box corresponding to the sensitive target within each image frame, determined by a state prediction module. Based on a pre-set weight adjustment strategy, the system determines the first weight corresponding to the first detection box, the second weight corresponding to the second detection box, the third weight corresponding to the third detection box, and the fourth weight corresponding to the fourth detection box. The system then performs weighted fusion of the first, second, third, and fourth detection boxes according to the first, second, third, and fourth weights to output the final detection box corresponding to the sensitive target within each image frame. Finally, the system locates the sensitive target region in the vehicle image based on the final detection box and performs desensitization processing on the sensitive target region based on a pre-set desensitization method.

[0010] According to another aspect of the present disclosure, a storage medium is also provided, the storage medium including a stored program, wherein, when the program is executed, a processor performs any of the methods described above.

[0011] According to another aspect of the present disclosure, a vehicle-mounted sensing sensitive target desensitization device is also provided, comprising: a sensitive target sensing module, configured to input acquired vehicle-mounted images into a pre-constructed information sensing model for sensitive target sensing, wherein the information sensing model includes a target detection network for detecting sensitive targets, a target tracking module for recording the trajectory of sensitive targets, an optical flow motion estimation module for estimating the cross-frame motion trend of sensitive targets, and a state prediction module for predicting the future motion state of sensitive targets, and the vehicle-mounted images include multiple image frames; and a detection box determination module, configured to determine a first detection box corresponding to the sensitive target in each image frame based on the target detection network, determine a second detection box corresponding to the sensitive target in each image frame based on the target tracking module, and determine the second detection box corresponding to the sensitive target in each image frame based on the optical flow motion estimation module. The system includes a third detection box corresponding to the sensitive target within each image frame, and a fourth detection box corresponding to the sensitive target within each image frame determined by a state prediction module; a weight determination module, used to determine the first weight corresponding to the first detection box, the second weight corresponding to the second detection box, the third weight corresponding to the third detection box, and the fourth weight corresponding to the fourth detection box based on a pre-set weight adjustment strategy; a final detection box output module, used to perform weighted fusion of the first, second, third, and fourth detection boxes according to the first, second, third, and fourth weights, and output the final detection box corresponding to the sensitive target within each image frame; and a desensitization module, used to locate the sensitive target region in the vehicle image based on the final detection box, and perform desensitization processing on the sensitive target region based on a preset desensitization method.

[0012] According to another aspect of the present disclosure, a vehicle-mounted sensing sensitive target desensitization device is also provided, comprising: a processor; and a memory connected to the processor, configured to provide the processor with instructions for processing the following steps: inputting acquired vehicle-mounted images into a pre-built information perception model for sensitive target perception, wherein the information perception model includes a target detection network for detecting sensitive targets, a target tracking module for recording the trajectory of sensitive targets, an optical flow motion estimation module for estimating the cross-frame motion trend of sensitive targets, and a state prediction module for predicting the future motion state of sensitive targets, and the vehicle-mounted images include multiple image frames; determining a first detection box corresponding to a sensitive target within each image frame based on the target detection network, and determining a second detection box corresponding to a sensitive target within each image frame based on the target tracking module. The system employs a two-box detection approach. A third detection box corresponding to the sensitive target within each image frame is determined based on an optical flow motion estimation module, and a fourth detection box corresponding to the sensitive target within each image frame is determined based on a state prediction module. Based on a pre-set weighting adjustment strategy, a first weight corresponding to the first detection box, a second weight corresponding to the second detection box, a third weight corresponding to the third detection box, and a fourth weight corresponding to the fourth detection box are determined. The first, second, third, and fourth detection boxes are weighted and fused according to their respective weights to output the final detection box corresponding to the sensitive target within each image frame. Finally, the system locates the sensitive target region in the vehicle-end image based on the final detection box and performs desensitization processing on the sensitive target region based on a pre-set desensitization method.

[0013] This application constructs an information perception model for sensing sensitive targets. This model includes a target detection network for detecting sensitive targets, a target tracking module for recording the trajectory of sensitive targets, an optical flow motion estimation module for estimating the cross-frame motion trend of sensitive targets, and a state prediction module for predicting the future motion state of sensitive targets. Therefore, this application can, after acquiring vehicle-side images, sequentially input each image frame of the vehicle-side image into the information perception model for sensitive target perception.

[0014] Then, the target detection network in the information perception model performs single-frame target detection on the current image frame and outputs all detection boxes corresponding to sensitive targets within the current image frame as the first detection boxes. Next, the target tracking module in the information perception model outputs the tracking boxes of each sensitive target within the current image frame as the second detection boxes. Simultaneously, the optical flow motion estimation module in the information perception model calculates the predicted positions of sensitive targets in the previous image frame within the current image frame and outputs the optical flow estimation boxes of each sensitive target within the current image frame as the third detection boxes. Subsequently, the state prediction module in the information perception model predicts the positions of each sensitive target within the current image frame and outputs the state prediction boxes of each sensitive target within the current image frame as the fourth detection boxes.

[0015] Furthermore, based on a pre-set weight adjustment strategy, a first weight corresponding to the first detection box, a second weight corresponding to the second detection box, a third weight corresponding to the third detection box, and a fourth weight corresponding to the fourth detection box are determined. Based on the first, second, third, and fourth weights, adaptive weighted fusion is performed on the first, second, third, and fourth detection boxes corresponding to the same sensitive target within the current image frame, outputting the final detection boxes for each sensitive target within the current image frame.

[0016] Finally, based on the coordinate information of the final detection box, the sensitive target region within each image frame is located, and a preset desensitization method is applied to the sensitive target region for local processing, while the original image data of the non-sensitive target region remains unchanged.

[0017] Therefore, this application introduces a target tracking module, an optical flow motion estimation module, and a state prediction module on the basis of the target detection network to construct an information perception model for sensing sensitive targets in vehicle-mounted images. Furthermore, this application utilizes a pre-set weight adjustment strategy to dynamically adjust the weights of each module based on the detection state, trajectory stability, motion continuity, and prediction error of the current image frame. Simultaneously, this application performs weighted fusion of the detection boxes corresponding to each module with the weights, and uses the weighted fused detection boxes as the final detection result. Thus, this application can achieve accurate perception and stable localization of sensitive targets in different vehicle-mounted scenarios. Therefore, this application can effectively suppress detection box jitter, reduce the target false negative rate, and improve localization stability. This solves the technical problems of insufficient robustness and unstable detection and localization in existing target detection methods for complex vehicle-mounted environments. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this disclosure and form part of this application, illustrate exemplary embodiments of this disclosure and are used to explain this disclosure, but do not constitute an undue limitation of this disclosure. In the drawings: Figure 1 This is a hardware structure block diagram of a computing device for implementing the method described in Embodiment 1 of this disclosure; Figure 2 This is a flowchart illustrating the vehicle-side sensing sensitive target desensitization method according to Embodiment 1 of this disclosure; Figure 3 This is a schematic diagram of the structure of the information perception model according to Embodiment 1 of this disclosure; Figure 4 This is a schematic diagram of the structure of the target detection network according to Embodiment 1 of this disclosure; Figure 5This is a flowchart illustrating the encryption and desensitization algorithm based on a chaotic system according to Embodiment 1 of this disclosure; Figure 6 This is a schematic diagram of the vehicle-end sensing sensitive target desensitization device according to Embodiment 2 of this disclosure; Figure 7 This is a schematic diagram of the vehicle-side sensing sensitive target desensitization device according to Embodiment 3 of this disclosure. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this disclosure.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] Example 1

[0022] According to this embodiment, a method for desensitizing vehicle-mounted sensing sensitive targets is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0023] The method embodiments provided in this example can be executed on mobile terminals, computer terminals, servers, or similar computing devices. Figure 1 A hardware block diagram of a computing device for implementing a vehicle-side sensitive target desensitization method is shown. Figure 1As shown, a computing device may include one or more processors (processors may include, but are not limited to, microprocessors such as MCUs or programmable logic devices such as FPGAs), a memory for storing data, a transmission device for communication functions, and an input / output interface. The memory, transmission device, and input / output interface are connected to the processor via a bus. In addition, it may also include a display, keyboard, and cursor control device connected to the input / output interface. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, a computing device may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0024] It should be noted that the aforementioned one or more processors and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element in a computing device. As involved in the embodiments of this disclosure, the data processing circuits serve as processor control (e.g., selection of a variable resistor termination path connected to an interface).

[0025] The memory can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the vehicle-side sensing sensitive target desensitization method in this embodiment of the present disclosure. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby implementing the aforementioned vehicle-side sensing sensitive target desensitization method of the application. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the computing device via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0026] The transmission device is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the computing device's communication provider. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0027] The display can be, for example, a touchscreen liquid crystal display (LCD), which allows users to interact with the user interface of the computing device.

[0028] It should be noted here that, in some optional embodiments, the above... Figure 1 The computing device shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 1 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computing devices.

[0029] Under the aforementioned operating environment, according to the first aspect of this embodiment, a method for desensitizing vehicle-side sensing sensitive targets is provided. This method comprises... Figure 1 The computing device shown is implemented. Figure 2 A flowchart illustrating the method is shown below. (Refer to...) Figure 2 As shown, the method includes: S202: The acquired vehicle-side image is input into a pre-built information perception model for sensitive target perception. The information perception model includes a target detection network for detecting sensitive targets, a target tracking module for recording the trajectory of sensitive targets, an optical flow motion estimation module for estimating the cross-frame motion trend of sensitive targets, and a state prediction module for predicting the future motion state of sensitive targets. The vehicle-side image includes multiple image frames. S204: Based on the target detection network, determine the first detection box corresponding to the sensitive target in each image frame; based on the target tracking module, determine the second detection box corresponding to the sensitive target in each image frame; based on the optical flow motion estimation module, determine the third detection box corresponding to the sensitive target in each image frame; and based on the state prediction module, determine the fourth detection box corresponding to the sensitive target in each image frame. S206: Based on a pre-set weight adjustment strategy, determine the first weight corresponding to the first detection box, the second weight corresponding to the second detection box, the third weight corresponding to the third detection box, and the fourth weight corresponding to the fourth detection box. S208: Based on the first weight, the second weight, the third weight, and the fourth weight, perform weighted fusion on the first detection box, the second detection box, the third detection box, and the fourth detection box, and output the final detection box corresponding to the sensitive target in each image frame; S210: Locate the sensitive target area in the vehicle image based on the final detection box, and perform desensitization processing on the sensitive target area based on the preset desensitization method.

[0030] Specifically, this application addresses the issues that existing improved YOLOv11 target detection networks still suffer from in complex vehicle-side dynamic scenarios, such as target omissions, detection box jitter, and unstable positioning across consecutive frames. Based on the improved YOLOv11 target detection network, this application further introduces a target tracking module for recording the trajectory of sensitive targets, an optical flow motion estimation module for estimating the cross-frame motion trend of sensitive targets, and a state prediction module for predicting the future motion state of sensitive targets to construct an information perception model for sensing sensitive targets.

[0031] In other words, reference Figure 3 As shown, after acquiring the vehicle-side image, the computing device sequentially inputs each image frame of the vehicle-side image into the information perception model for sensitive target perception (corresponding to step S202).

[0032] Next, the computing device inputs the current image frame into the target detection network. After performing a single-frame target detection task on the current image frame through the target detection network, the detection result is output, which includes the first detection box, category label, and confidence score.

[0033] The first detection box is a rectangle, and the computing device represents the first detection box of each detected sensitive target in the form of a quadruple. The quadruple is (x_center, y_center, width, height), where x_center represents the x-coordinate of the center point of the first detection box, y_center represents the y-coordinate of the center point of the first detection box, width represents the width of the first detection box, and height represents the height of the first detection box. The computing device normalizes the coordinate values ​​of the first detection box to the interval [0,1].

[0034] The category label is used to indicate the target category to which the detected sensitive target belongs, and the target category includes face category and license plate category.

[0035] The confidence score ranges from [0,1] and is used to represent the probability that the target detection network determines that the region enclosed by the first detection box contains the target of the corresponding category.

[0036] Furthermore, the computing device uses the target tracking module to perform inter-frame correlation matching between the first detection box in the current image frame and the target trajectories recorded in previous image frames to determine the historical target category corresponding to each sensitive target in the current image frame. Simultaneously, the computing device outputs the tracking boxes of each sensitive target in the current image frame based on the matching results, serving as the second detection boxes.

[0037] Meanwhile, the computing device uses the optical flow motion estimation module to calculate the pixel-level motion vector field between adjacent image frames, infers the predicted position of the sensitive target in the previous image frame in the current image frame based on the motion vector field, and outputs the optical flow estimation box of each sensitive target in the current image frame as the third detection box.

[0038] Then, the computing device uses the state prediction module to predict the position of each sensitive target in the current image frame and outputs the state prediction box of each sensitive target in the current image frame as the fourth detection box (corresponding to step S204).

[0039] Furthermore, the computing device, based on a pre-set weight adjustment strategy, determines the values ​​relative to the first detection frame. The corresponding first weight and the second detection frame The corresponding second weight and the third detection frame The corresponding third weight and the fourth detection box The corresponding fourth weight (Corresponding to step S206).

[0040] The computing device can dynamically adjust the first weight based on the environmental state, trajectory stability, motion continuity, and prediction error of the current image frame. Second weight Third weight and the fourth weight And satisfy: (1) Then, for each target trajectory, the computing device calculates according to the first weight. Second weight Third weight and the fourth weight The first detection box corresponding to the same sensitive target within the current image frame. Second detection frame Third detection frame and the fourth detection frame Perform adaptive weighted fusion to output the final detection bounding boxes for each sensitive target within the current image frame. (Corresponding to step S208). The calculation formula is as follows: (2) Next, the computing device determines the image region enclosed by the final detection box within the current image frame of the vehicle-end image based on the coordinate information of the final detection box, and marks this image region as a sensitive target region. Then, the computing device performs desensitization operation on the sensitive target region according to a preset desensitization method, while keeping the original image data of the non-sensitive target regions in the current image frame unchanged (corresponding to step S210).

[0041] As described in the background section, most existing detection methods employ single-frame target detection, relying primarily on the target detection network to directly output the detection results, lacking effective utilization of video temporal information. In real-world vehicle scenarios, due to complex factors such as high-speed vehicle movement, short-term target occlusion, low brightness at night, and motion blur, simply relying on the detection model can easily lead to problems such as detection box jitter, missed target detection, and unstable localization.

[0042] Existing vehicle-side desensitization technologies generally focus on the desensitization process itself, employing a single method to handle sensitive targets. This not only lacks flexible configuration capabilities for different application scenarios but also fails to provide effective quantitative evaluation of the desensitization effect. In practical applications, simply completing the desensitization operation without assessing whether the desensitization is sufficient, whether there are omissions or over-processing, makes it difficult to meet the requirement of verifiability of the desensitization results.

[0043] In view of this, this application constructs an information perception model for sensing sensitive targets. This information perception model includes a target detection network for detecting sensitive targets, a target tracking module for recording the trajectory of sensitive targets, an optical flow motion estimation module for estimating the cross-frame motion trend of sensitive targets, and a state prediction module for predicting the future motion state of sensitive targets. Therefore, this application can, after acquiring vehicle-side images, sequentially input each image frame of the vehicle-side image into the information perception model for sensitive target perception.

[0044] Then, the target detection network in the information perception model performs single-frame target detection on the current image frame and outputs all detection boxes corresponding to sensitive targets within the current image frame as the first detection boxes. Next, the target tracking module in the information perception model outputs the tracking boxes of each sensitive target within the current image frame as the second detection boxes. Simultaneously, the optical flow motion estimation module in the information perception model calculates the predicted positions of sensitive targets in the previous image frame within the current image frame and outputs the optical flow estimation boxes of each sensitive target within the current image frame as the third detection boxes. Subsequently, the state prediction module in the information perception model predicts the positions of each sensitive target within the current image frame and outputs the state prediction boxes of each sensitive target within the current image frame as the fourth detection boxes.

[0045] Furthermore, based on a pre-set weight adjustment strategy, a first weight corresponding to the first detection box, a second weight corresponding to the second detection box, a third weight corresponding to the third detection box, and a fourth weight corresponding to the fourth detection box are determined. Based on the first, second, third, and fourth weights, adaptive weighted fusion is performed on the first, second, third, and fourth detection boxes corresponding to the same sensitive target within the current image frame, outputting the final detection boxes for each sensitive target within the current image frame.

[0046] Finally, based on the coordinate information of the final detection box, the sensitive target region within each image frame is located, and a preset desensitization method is applied to the sensitive target region for local processing, while the original image data of the non-sensitive target region remains unchanged.

[0047] Therefore, this application introduces a target tracking module, an optical flow motion estimation module, and a state prediction module on the basis of the target detection network to construct an information perception model for sensing sensitive targets in vehicle-mounted images. Furthermore, this application utilizes a pre-set weight adjustment strategy to dynamically adjust the weights of each module based on the detection state, trajectory stability, motion continuity, and prediction error of the current image frame. Simultaneously, this application performs weighted fusion of the detection boxes corresponding to each module with the weights, and uses the weighted fused detection boxes as the final detection result. Thus, this application can achieve accurate perception and stable localization of sensitive targets in different vehicle-mounted scenarios. Therefore, this application can effectively suppress detection box jitter, reduce the target false negative rate, and improve localization stability. This solves the technical problems of insufficient robustness and unstable detection and localization in existing target detection methods for complex vehicle-mounted environments.

[0048] Furthermore, to cover the actual needs of different application scenarios in vehicles, the computing device can acquire vehicle images by uploading pre-recorded complete video files from the cloud, or by capturing video streams of relevant vehicle images in real time through devices such as vehicle cameras and dashcams. Subsequently, the computing device decodes and resizes the acquired vehicle images frame by frame, converting them into a unified input format to meet the input requirements of subsequent information perception models.

[0049] One method involves the computing device uploading a fully recorded video file from the cloud to acquire vehicle-side images. This approach is suitable for post-processing scenarios where dashcams and similar devices have already recorded video. The computing device then receives the video file to be processed by the user and stores it in a designated directory. This method supports common video encoding formats such as H.264 and H.265, as well as video container formats such as MP4, AVI, and MOV. After the user uploads the video file, the computing device automatically parses the video file's metadata, including parameters such as video duration, resolution, and frame rate, preparing for subsequent frame-by-frame processing of the vehicle-side images.

[0050] Meanwhile, the computing device acquires video streams of vehicle-side images in real time through devices such as vehicle-mounted cameras and dashcams, making it suitable for real-time acquisition and processing scenarios during driving. The computing device establishes a real-time video stream connection by calling the application programming interface (API) of the vehicle-mounted camera or dashcam camera module. The video stream is accessed via the RTSP protocol or by direct capture from the local camera. During real-time acquisition, the camera transmits each captured image frame to the computing device immediately, enabling a pipelined workflow of simultaneous acquisition and processing.

[0051] Then, the computing device decodes and preprocesses the acquired vehicle-side images frame by frame to meet the input requirements of the subsequent information perception model. Specific preprocessing operations include size adjustment, color space conversion, and dimensional expansion. After preprocessing, a single frame image is expanded from a three-dimensional array of (height, width, channels) to a four-dimensional tensor of (batch_size, channels, height, width), where batch_size is initially set to 1 to meet the batch processing dimensionality requirements during inference by the information perception model.

[0052] It should be noted that this embodiment uses an improved YOLOv11 object detection network as the basic detection module, enabling it to adapt to the low-computing-power operating environment of vehicles while maintaining high detection accuracy. This application uses the YOLOv11n version as the baseline model for improvement. This version has the fewest parameters and the fastest inference speed in the YOLOv11 series, making it suitable as a starting point for improvements for vehicle-embedded environments.

[0053] like Figure 4 As shown, the improved YOLOv11 target detection network consists of three parts: Backbone, Neck, and Head.

[0054] In the backbone, the computing device takes the preprocessed image frames as input, which are then processed sequentially through a lightweight convolutional module (GConv), a feature extraction module (C3k2 module), and a multi-scale feature fusion module (SPPF module) for feature extraction before entering the attention enhancement module (C2PSA). Notably, the computing device replaces some of the convolutional modules (Conv) surrounding the C3k2 module with lightweight convolutional modules (GConv) to reduce the number of model parameters and computational complexity. Simultaneously, the computing device embeds a parameterless attention module (SimAM) in the backbone, which calculates feature importance based on neuron energy functions, enhancing the model's responsiveness to sensitive target regions such as faces and license plates without adding additional parameters.

[0055] In the Neck section, the computing device performs feature fusion on the multi-scale feature maps output by the Backbone through the convolution module (Conv), upsampling module (Upsample), concatenation module (Concat), and C3k2 module. This aggregates semantic and positional information from different scales to improve the model's ability to detect sensitive targets of different sizes.

[0056] In the Head section, the computing device performs target detection on the fused feature map output by the Neck through the detection layer, and outputs the position of the first detection box, category label and confidence score of each sensitive target in the current image frame.

[0057] It should be noted that, in Figure 4 In the diagram, dark-filled modules (such as lightweight convolutional modules and parameterless attention modules) represent the improvements made to the standard YOLOv11 object detection network compared to the standard YOLOv11 network, while the remaining modules are the original modules in the standard YOLOv11 network. Through the above structural improvements, the improved YOLOv11 object detection network significantly reduces computational resource consumption while maintaining high detection accuracy, making it suitable for low-computing-power operating environments on vehicle devices.

[0058] Optionally, the weight adjustment strategy includes: associating a first weight with the confidence score output by the target detection network; associating a second weight with the duration of the motion trajectory of the sensitive target; associating a third weight with the motion consistency of the sensitive target; and associating a fourth weight with the prediction error generated by the state prediction module.

[0059] Specifically, the computing device can dynamically adjust the first weight, second weight, third weight, and fourth weight according to a pre-set weight adjustment strategy.

[0060] Specifically, the computing device correlates a first weight with a confidence score output by the object detection network. This confidence score reflects the reliability of the object detection network in confirming that its first detection box contains a sensitive target. A higher confidence score indicates higher reliability of the first detection box. Therefore, the computing device can adjust the first weight based on the confidence score; a higher confidence score results in a higher first weight.

[0061] Simultaneously, the computing device correlates the second weight with the duration of the sensitive target's trajectory. The duration of the trajectory reflects the length of time the target tracking module continuously tracks the same sensitive target. A longer trajectory duration indicates higher trajectory stability and greater reliability of the tracking results. Therefore, the computing device can determine the second weight based on the duration of the sensitive target's trajectory; the longer the trajectory duration, the higher the second weight.

[0062] Furthermore, the computing device correlates the third weight with the motion consistency of the sensitive target. This motion consistency reflects the degree of directional change when the optical flow motion estimation module infers the motion direction of the sensitive target between consecutive image frames. Higher motion consistency indicates a stable motion direction between consecutive image frames, while lower motion consistency indicates a drastic change in motion direction. Therefore, the computing device can determine the third weight based on motion consistency; higher motion consistency results in a higher third weight.

[0063] The computing device correlates the fourth weight with the prediction error generated by the state prediction module. This prediction error reflects the degree of deviation between the state prediction box output by the state prediction module and the actual position of the sensitive target. A smaller prediction error indicates that the estimation result of the state prediction module is more accurate, while a larger prediction error indicates that the estimation result is less reliable. Therefore, the computing device determines the fourth weight based on the prediction error; the smaller the prediction error, the higher the fourth weight, and the larger the prediction error, the lower the fourth weight.

[0064] In this way, the contribution of each module within the current image frame is matched with the reliability of its output.

[0065] Optionally, the weight adjustment strategy further includes: classifying the current vehicle scene into static scenes, high-motion scenes, nighttime scenes, and blurred scenes based on the environmental features included in the vehicle-end image; in static scenes, increasing the first weight of the target detection network and decreasing the third weight of the optical flow motion estimation module; in high-motion scenes, increasing the second weight of the target tracking module and the third weight of the optical flow motion estimation module; in nighttime scenes, increasing the second weight of the target tracking module and the fourth weight of the state prediction module; and in blurred scenes, decreasing the first weight of the target detection network and increasing the second weight of the target tracking module and the fourth weight of the state prediction module.

[0066] Specifically, the computing device further classifies the current vehicle-end image into static scenes, high-motion scenes, nighttime scenes, and blurred scenes based on the environmental characteristics reflected in the current vehicle-end image. Then, the computing device dynamically adjusts the first, second, third, and fourth weights according to the classified scene type.

[0067] Specifically, because the object detection network has high detection accuracy in static scenes, its output first detection box has a high confidence score. Therefore, the computing device will increase the first weight of the object detection network in static scenes. At the same time, the reliability of the optical flow motion estimation module's calculation results is low in static scenes, so the computing device will decrease the third weight of the optical flow motion estimation module.

[0068] However, because the target tracking module is based on trajectory matching of historical image frames, it can maintain relatively stable tracking output in high-motion scenarios where the vehicle is traveling at high speed. Therefore, the computing device will increase the second weight corresponding to the second detection box output by the target tracking module in high-motion scenarios. At the same time, the optical flow motion estimation module can still reflect the displacement trend of sensitive targets well when the target motion intensity is large. Therefore, the computing device will increase the third weight corresponding to the optical flow motion estimation module in high-motion scenarios.

[0069] Similarly, in nighttime scenes, the ambient light is low, and the image signal-to-noise ratio decreases. However, the target tracking module, relying on the continuity of historical motion trajectories, can still provide relatively stable tracking boxes. Meanwhile, the state prediction module, based on a Kalman filter motion model, does not depend on the image quality of the current frame and can still output reliable state prediction boxes in nighttime scenes. Therefore, the computing device increases the second weight corresponding to the target tracking module and the fourth weight corresponding to the state prediction module in nighttime scenes.

[0070] In blurred scenes, vehicle-side images suffer from detail loss due to motion blur or refocusing issues, significantly reducing the detection accuracy of the object detection network. Specifically, the first bounding box is prone to location shifts and a drop in confidence score. Therefore, the computing device reduces the first weight corresponding to the object detection network in blurred scenes. However, the object tracking module and the state prediction module do not depend on the image sharpness of the current frame and are less affected by image quality degradation in blurred scenes. Therefore, the computing device increases the second weight corresponding to the object tracking module and the fourth weight corresponding to the state prediction module in blurred scenes.

[0071] In this way, the computing device can flexibly adjust the weights of each module in different vehicle operating environments, so that the information perception model can maintain high sensitivity target perception accuracy and positioning stability under different scenario conditions.

[0072] Optionally, the environmental features include brightness features, sharpness features, global motion features, and sensitive target motion features. The operation of classifying the current vehicle scene into static scenes, high-motion scenes, nighttime scenes, and blurred scenes based on the environmental features included in the vehicle-end image includes: extracting the brightness features, sharpness features, global motion features, and sensitive target motion features of the current image frame; determining the current vehicle scene as a nighttime scene when the ambient brightness corresponding to the brightness feature is lower than a first preset threshold; determining the current vehicle scene as a blurred scene when the image sharpness corresponding to the sharpness feature is lower than a second preset threshold; determining the current vehicle scene as a high-motion scene when the inter-frame motion intensity corresponding to the global motion feature is higher than a third preset threshold; and determining the current vehicle scene as a static scene when the inter-frame motion intensity is lower than the third preset threshold and the target motion intensity corresponding to the sensitive target motion feature is lower than a fourth preset threshold.

[0073] Specifically, the computing device categorizes the current vehicle scene into static scenes, high-motion scenes, nighttime scenes, and blurred scenes based on the extracted brightness features, sharpness features, global motion features, and sensitive target motion features of the current image frame. Brightness features are used to determine whether the environment reflected in the current image frame is a low-light nighttime condition; sharpness features are used to determine whether the current image frame is blurred due to motion or focus issues; global motion features are used to determine the motion intensity of the vehicle itself or the overall image; and sensitive target motion features are used to determine the motion state of a single sensitive target.

[0074] Specifically, the computing device can calculate the average brightness of the grayscale image pixels corresponding to the current image frame using the following formula to determine the ambient brightness corresponding to the brightness feature: (3) Where N is the total number of pixels in the grayscale image I corresponding to the current image frame. It represents the value of the i-th pixel in the grayscale image I corresponding to the current image frame.

[0075] When the ambient brightness corresponding to the brightness feature is lower than the first preset threshold, the computing device determines that the current vehicle scene is a night scene.

[0076] Furthermore, computing devices can measure image sharpness using the Laplacian variance value Blur using the following formula: (4) in, This is used to represent the result of performing a Laplacian convolution operation on the grayscale image I corresponding to the current image frame. (·) indicates variance operation.

[0077] The larger the Laplacian variance value Blur, the higher the image sharpness of the current image frame. When the image sharpness corresponding to the sharpness feature is lower than the second preset threshold, the computing device determines that the current vehicle-side scene is a blurred scene.

[0078] Furthermore, the computing device can calculate the GlobalMotion intensity between consecutive frames using the dense optical flow algorithm, namely: (5) Where u represents the horizontal component of the optical flow field, and v represents the vertical component. The optical flow field is a two-dimensional vector field calculated from two consecutive image frames. Each pixel corresponds to a motion vector, which is composed of the horizontal component u and the vertical component v, representing the displacement or velocity components of the pixel from the previous image frame to the current image frame in the horizontal and vertical directions, respectively; median(·) represents the median operation.

[0079] When the intensity of continuous inter-frame motion corresponding to global motion features is higher than the third preset threshold, the computing device determines that the current vehicle-side scene is a high-motion scene. Additionally, the computing device can estimate the target motion intensity, VehicleMotion, based on the displacement of the vehicle's center point within consecutive image frames using the following formula: (6) This represents the x-coordinate of the center point of the same sensitive target in the current image frame. This represents the ordinate of the center point of the same sensitive target in the current image frame.

[0080] This represents the x-coordinate of the center point of the same sensitive target within the previous image frame; This represents the ordinate of the center point of the same sensitive target within the previous image frame.

[0081] When the motion intensity between consecutive frames is lower than the third preset threshold and the motion intensity of the target corresponding to the motion feature of the sensitive target is lower than the fourth preset threshold, the computing device determines that the current vehicle-side scene is a static scene.

[0082] It should be noted that global motion features reflect the overall motion intensity of the image background (mainly determined by the vehicle's own speed), while sensitive target motion features reflect the inter-frame displacement intensity of the sensitive target itself in the image. They play different roles in vehicle-side scene segmentation. For example, when the vehicle is stationary but pedestrians or non-motorized vehicles are moving rapidly in the image, global motion features are lower, but sensitive target motion features are higher.

[0083] Thus, through the above methods, the computing device can classify the current vehicle scene into static scenes, high-motion scenes, nighttime scenes, and blurred scenes based on the extracted environmental features.

[0084] Optionally, when the detection of a sensitive target fails, the information perception model uses the third detection box, the first recovery weight coefficient corresponding to the third detection box, the fourth detection box, and the second recovery weight coefficient corresponding to the fourth detection box to perform secondary fusion on the current image frame of the vehicle-end image, and uses the fused detection box as the final detection box of the sensitive target in the current image frame.

[0085] Specifically, when the target detection network in the information perception model fails to output a valid first detection box for a sensitive target in the current image frame (i.e., due to complex factors such as short-term occlusion of the target, motion blur, or low light at night), or when the confidence score of the first detection box is lower than a preset threshold, the computing device determines that the sensitive target detection has failed.

[0086] In this situation, the computing device activates the missed detection recovery mechanism to perform secondary fusion on the current image frame of the vehicle-side image.

[0087] Specifically, the computing device first acquires the third detection box output by the optical flow motion estimation module and the fourth detection box output by the state prediction module. Then, the computing device assigns a first recovery weight coefficient to the third detection box. Assign a second recovery weight coefficient to the fourth detection box. And it satisfies the following formula: + =1 (7) In scenarios where detection fails, the target detection network cannot perform target detection on the current image frame and can only rely on the third and fourth detection boxes to obtain target information. Therefore, the computing device performs secondary fusion on the current image frame of the vehicle-end image based on the third detection box, the first restoration weight coefficient corresponding to the third detection box, the fourth detection box, and the second restoration weight coefficient corresponding to the fourth detection box, and outputs the fused detection boxes as the final detection boxes (Brec) for sensitive targets within the current image frame. That is: (8) Thus, through the above-mentioned missed detection recovery mechanism, the computing device can continuously provide position estimates for sensitive targets by relying on the collaborative calculation of the optical flow motion estimation module and the state prediction module when the target detection network fails to detect.

[0088] Optionally, the operation of locating the sensitive target region in the vehicle-end image based on the final detection box and performing desensitization processing on the sensitive target region based on a preset desensitization method includes: determining the image region enclosed by the final detection box within the current image frame of the vehicle-end image based on the coordinate information of the final detection box, and marking the image region as a sensitive target region; and performing desensitization operation on the sensitive target region according to the preset desensitization method, while keeping the original image data of the non-sensitive target regions within the current image frame unchanged.

[0089] Specifically, the computing device first obtains the coordinate information of the final detection bounding boxes of each sensitive target within the current image frame. This coordinate information is represented as a quadruple (x_center, y_center, width, height). Then, based on this coordinate information, the computing device determines the rectangular image region enclosed by the final detection bounding boxes in the pixel coordinate system of the current image frame, thereby locating the specific position of this rectangular image region within the current image frame. Subsequently, the computing device labels the determined rectangular image region as the sensitive target region.

[0090] Furthermore, the computing device performs desensitization operations only on the identified sensitive target areas according to the user's pre-set desensitization method. In this embodiment, the preset desensitization methods include traditional visual desensitization methods (such as Gaussian blur, mosaic processing, and pixelation processing) and reversible encryption desensitization methods based on the SLCS chaotic system.

[0091] It should be noted that, regardless of the desensitization method used, the computing device only processes the pixel data within the sensitive target area, and does not perform desensitization operations on all non-sensitive target areas within the current image frame except for the sensitive target area, so as to keep its original image data unchanged.

[0092] Therefore, compared with the existing technology that desensitizes the entire image frame, the operation of desensitizing a local area of ​​the current image frame significantly reduces the number of pixels that need to be processed, and has higher processing efficiency under the condition of limited computing power on the vehicle.

[0093] Optionally, the sensitive target area is desensitized according to a preset desensitization method, including: receiving a string key input by the user, generating a fixed-length hash value using a hash algorithm, and mapping the hash result to encryption control parameters, including the initial value and control parameters of the SLCS chaotic system and diffusion and scrambling control parameters; generating a key stream using the SLCS chaotic system based on the initial value and control parameters; constructing a cryptographic matrix based on the key stream and parameterizing the cryptographic matrix according to the diffusion and scrambling control parameters; performing a forward diffusion operation on the sensitive target area to couple pixel values ​​with neighboring pixels and the cryptographic matrix; generating a dynamic scrambling index using the cryptographic matrix and performing a non-linear scrambling operation on the diffused pixel sequence; and performing a backward diffusion operation on the scrambling result to convert the sensitive target area into an unrecognizable ciphertext image area and replace the corresponding area in the original vehicle-end image.

[0094] Specifically, this embodiment designs a method for desensitizing the sensitive target area according to a preset desensitization method. This method can be applied to high-security application scenarios. Unlike traditional visual desensitization methods that encrypt the entire image frame, this application only performs local encryption processing on the detected sensitive target areas, including the face area and license plate area, thereby significantly reducing computational overhead and improving vehicle-side operating efficiency.

[0095] Specifically, refer to Figure 5 As shown, the computing device first receives the string key input by the user and generates a fixed-length hash value using the SHA-256 hash algorithm. The hash result is then mapped to eight cryptographic control parameters:

[0096] in , , , represent the initial state values ​​of the two chaotic sequences used in the iterative calculation of the SLCS chaotic system. p and q are the two chaotic control parameters of the SLCS chaotic system, respectively. , , , The control parameters represent diffusion and disorder.

[0097] It should be noted that this embodiment uses an improved SLCS chaotic system, commonly used in image encryption, to generate a highly random keystream, thereby improving the randomness, key sensitivity, and resistance to attacks during the encryption process. The iterative formula of the SLCS chaotic system is defined as follows: (9) in: ∈(0,1) represents the state value of the nth iteration, p represents the chaotic control parameter, and n represents the iteration number. To ensure that the output of the SLCS chaotic system is always within the effective interval, it is normalized after each iteration according to the following formula to map it to the interval [0,1).

[0098] (10) In actual calculations, this embodiment generates two independent chaotic sequences simultaneously: (11) (12) To eliminate the influence of the initial transition state on the chaotic properties, the SLCS chaotic system first skips a preset number of initial iteration results, and then extracts the chaotic sequence in the stable state for subsequent encryption calculations.

[0099] After obtaining the chaotic sequence, the SLCS chaotic system further constructs four cryptographic matrices X, Y, R, and W. X is used for forward diffusion, Y for backward diffusion, and R and W for dynamic scrambling.

[0100] Furthermore, the SLCS chaotic system first performs a forward diffusion operation on the image frame of the sensitive target region, so that the pixel value is coupled with the neighboring pixels and the cryptographic matrix X, thereby enhancing the sensitivity of the ciphertext to small changes in the plaintext.

[0101] Meanwhile, the SLCS chaotic system uses the cryptographic matrix R and cryptographic matrix W to generate a dynamic scrambling index, performing a nonlinear scrambling operation on the diffused pixel sequence. Furthermore, this embodiment introduces a cyclic shift mechanism, performing a cyclic left shift operation on pixels before swapping their positions, thereby enhancing the nonlinear characteristics of the encryption process and improving its resistance to statistical attacks.

[0102] In addition, the SLCS chaotic system performs backdiffusion on the scrambling result, so that the current pixel is simultaneously affected by the subsequent pixels and the cryptographic matrix Y, further enhancing the randomness of the ciphertext.

[0103] After the above processing, the sensitive target area is converted into an unrecognizable encrypted image area and replaces the corresponding area in the original image frame.

[0104] In addition, this application also supports authorized recovery of encrypted areas. Users with legitimate keys can perform reverse backdiffusion, reverse dynamic scrambling, and reverse forward diffusion operations to fully decrypt and recover encrypted ciphertext image areas.

[0105] It should be noted that, in order to fill the gap in the existing technology of lacking a quantitative evaluation method for desensitization effect, this application constructs a multi-dimensional desensitization effect evaluation index system to systematically evaluate the effect of desensitized vehicle images.

[0106] The evaluation indicators and their definitions are as follows: (1) Coverage MR: used to measure the degree of occlusion of sensitive targets by desensitization operation.

[0107] The computing device averages the coverage of all image frames to obtain the overall coverage of the vehicle-side image. A higher overall coverage indicates a more comprehensive desensitization process. If we let... This represents the desensitization region at position f. The formula is as follows, indicating the corresponding area to be desensitized: (13) In the formula, m is the total number of desensitized regions.

[0108] (2) False positive rate and misdetection rate: These are used to construct a confusion matrix to test the classification performance of the information perception model. The false positive rate represents the percentage of false positives. A higher false positive rate indicates a greater tendency to desensitize non-sensitive targets, which can lead to unnecessary computational resource consumption. The misdetection rate quantifies the degree to which the information perception model misses sensitive targets. A higher misdetection rate makes it easier to overlook sensitive targets in real-world data, failing to effectively desensitize all sensitive targets and resulting in insufficient recognition capabilities to meet actual needs.

[0109] In this confusion matrix, rows are arranged by predicted class and columns by true class. A true positive (TP) is correctly classified as a positive; a false positive (FP) is incorrectly classified as a negative; a true negative (TN) is correctly classified as a negative; and a false negative (FN) is incorrectly classified as a positive.

[0110] It should be noted that, in the scenario of desensitizing sensitive information in vehicle-side images, TP indicates that the computing device has successfully identified and desensitized the sensitive information regions in the test set, and the intersection-union ratio (IoU) between the labeled sensitive target region and the actual desensitized region is greater than 80%. FP indicates that the computing device identified non-sensitive target regions in the test set as sensitive target regions and performed invalid desensitization, and the actual desensitized region has no intersection with any labeled sensitive target region. TN indicates that the computing device did not identify the non-sensitive target regions set in the test set as sensitive target regions; FN indicates that the computing device failed to identify the sensitive target regions in the test set, and the labeled sensitive target region has no intersection with any actual desensitized region.

[0111] Based on the above concepts, the computing device calculates the false positive rate (FPR) according to the following formula: (14) The computing device calculates the false positive rate (FNR) using the following formula: (15) Next, the computing device averages the false detection rates of all image frames in the vehicle-side image to obtain the overall false detection rate of the vehicle-side image. The lower the false detection rate, the lower the degree of omission in the de-identification process.

[0112] (3) Desensitization success rate: The target detection network is used to re-identify the desensitized image frame. If the sensitive target cannot be identified, the image frame is successfully desensitized; otherwise, if the target detection network can identify the sensitive target, the image frame fails to be desensitized.

[0113] The success rate of desensitization is the ratio of successfully desensitized frames to the total number of frames. The success rate comprehensively reflects the overall quality of the desensitization effect; a higher success rate indicates a better desensitization effect.

[0114] Furthermore, when evaluating the desensitization results, the computing device analyzes the desensitized vehicle-side images to be evaluated frame by frame, calculating the coverage rate, false detection rate, and misdetection rate for each frame. After the overall evaluation is completed, the computing device summarizes the index data of all image frames, calculates the comprehensive desensitization success rate of the vehicle-side images, and automatically generates an evaluation report. The evaluation report includes basic image information, an overall desensitization success rate score, and displays the specific frame number, thumbnail, and problem description of the unqualified image frames.

[0115] By employing the aforementioned multi-dimensional and quantifiable desensitization effectiveness evaluation methods, this application effectively addresses the shortcomings of existing desensitization schemes in effectiveness verification, providing a scientific and systematic evaluation method for the security and compliance of vehicle-side image data.

[0116] In addition, refer to Figure 1 As shown, according to a second aspect of this embodiment, a storage medium is provided. The storage medium includes a stored program, wherein, when the program is executed, a processor performs any of the methods described above.

[0117] Therefore, according to this embodiment, this application introduces a target tracking module, an optical flow motion estimation module, and a state prediction module on the basis of the target detection network to construct an information perception model for sensing sensitive targets in vehicle-side images. Furthermore, this application utilizes a pre-set weight adjustment strategy to dynamically adjust the weights of each module based on the detection state, trajectory stability, motion continuity, and prediction error of the current image frame. Simultaneously, this application performs weighted fusion of the detection boxes corresponding to each module with the weights, and uses the weighted fused detection boxes as the final detection result. Thus, this application can achieve accurate perception and stable localization of sensitive targets in different vehicle-side scenarios. Therefore, this application can effectively suppress detection box jitter, reduce the target false negative rate, and improve localization stability. This solves the technical problems of insufficient robustness and unstable detection and localization in existing target detection methods for complex vehicle-side environments.

[0118] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0119] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0120] Example 2

[0121] Figure 6 A vehicle-mounted sensing sensitive target desensitization device according to this embodiment is shown, which corresponds to the method described according to Embodiment 1. Reference Figure 6As shown, the device includes: a sensitive target perception module 610, used to input acquired vehicle-side images into a pre-built information perception model for sensitive target perception, wherein the information perception model includes a target detection network for detecting sensitive targets, a target tracking module for recording the trajectory of sensitive targets, an optical flow motion estimation module for estimating the cross-frame motion trend of sensitive targets, and a state prediction module for predicting the future motion state of sensitive targets, and the vehicle-side images include multiple image frames; a detection box determination module 620, used to determine a first detection box corresponding to the sensitive target in each image frame based on the target detection network, a second detection box corresponding to the sensitive target in each image frame based on the target tracking module, and a third detection box corresponding to the sensitive target in each image frame based on the optical flow motion estimation module; and... The state prediction module determines the fourth detection box corresponding to the sensitive target within each image frame; the weight determination module 630 is used to determine the first weight corresponding to the first detection box, the second weight corresponding to the second detection box, the third weight corresponding to the third detection box, and the fourth weight corresponding to the fourth detection box based on a preset weight adjustment strategy; the final detection box output module 640 is used to perform weighted fusion of the first detection box, the second detection box, the third detection box, and the fourth detection box according to the first weight, the second weight, the third weight, and the fourth weight, and output the final detection box corresponding to the sensitive target within each image frame; and the desensitization module 650 is used to locate the sensitive target region in the vehicle image according to the final detection box, and perform desensitization processing on the sensitive target region according to a preset desensitization method.

[0122] Optionally, the weight adjustment strategy includes: associating a first weight with the confidence score output by the target detection network; associating a second weight with the duration of the motion trajectory of the sensitive target; associating a third weight with the motion consistency of the sensitive target; and associating a fourth weight with the prediction error generated by the state prediction module.

[0123] Optionally, the weight adjustment strategy further includes: classifying the current vehicle scene into static scenes, high-motion scenes, nighttime scenes, and blurred scenes based on the environmental features included in the vehicle-end image; in static scenes, increasing the first weight of the target detection network and decreasing the third weight of the optical flow motion estimation module; in high-motion scenes, increasing the second weight of the target tracking module and the third weight of the optical flow motion estimation module; in nighttime scenes, increasing the second weight of the target tracking module and the fourth weight of the state prediction module; and in blurred scenes, decreasing the first weight of the target detection network and increasing the second weight of the target tracking module and the fourth weight of the state prediction module.

[0124] Optionally, the environmental features include brightness features, sharpness features, global motion features, and sensitive target motion features. The operation of classifying the current vehicle scene into static scenes, high-motion scenes, nighttime scenes, and blurred scenes based on the environmental features included in the vehicle-end image includes: extracting the brightness features, sharpness features, global motion features, and sensitive target motion features of the current image frame; determining the current vehicle scene as a nighttime scene when the ambient brightness corresponding to the brightness feature is lower than a first preset threshold; determining the current vehicle scene as a blurred scene when the image sharpness corresponding to the sharpness feature is lower than a second preset threshold; determining the current vehicle scene as a high-motion scene when the inter-frame motion intensity corresponding to the global motion feature is higher than a third preset threshold; and determining the current vehicle scene as a static scene when the inter-frame motion intensity is lower than the third preset threshold and the target motion intensity corresponding to the sensitive target motion feature is lower than a fourth preset threshold.

[0125] Optionally, when the detection of a sensitive target fails, the information perception model uses the third detection box, the first recovery weight coefficient corresponding to the third detection box, the fourth detection box, and the second recovery weight coefficient corresponding to the fourth detection box to perform secondary fusion on the current image frame of the vehicle-end image, and uses the fused detection box as the final detection box of the sensitive target in the current image frame.

[0126] Optionally, the operation of locating the sensitive target region in the vehicle-end image based on the final detection box and performing desensitization processing on the sensitive target region based on a preset desensitization method includes: determining the image region enclosed by the final detection box within the current image frame of the vehicle-end image based on the coordinate information of the final detection box, and marking the image region as a sensitive target region; and performing desensitization operation on the sensitive target region according to the preset desensitization method, while keeping the original image data of the non-sensitive target regions within the current image frame unchanged.

[0127] Optionally, the sensitive target area is desensitized according to a preset desensitization method, including: receiving a string key input by the user, generating a fixed-length hash value using a hash algorithm, and mapping the hash result to encryption control parameters, including the initial value and control parameters of the SLCS chaotic system and diffusion and scrambling control parameters; generating a key stream using the SLCS chaotic system based on the initial value and control parameters; constructing a cryptographic matrix based on the key stream and parameterizing the cryptographic matrix according to the diffusion and scrambling control parameters; performing a forward diffusion operation on the sensitive target area to couple pixel values ​​with neighboring pixels and the cryptographic matrix; generating a dynamic scrambling index using the cryptographic matrix and performing a non-linear scrambling operation on the diffused pixel sequence; and performing a backward diffusion operation on the scrambling result to convert the sensitive target area into an unrecognizable ciphertext image area and replace the corresponding area in the original vehicle-end image.

[0128] Therefore, according to this embodiment, this application introduces a target tracking module, an optical flow motion estimation module, and a state prediction module on the basis of the target detection network to construct an information perception model for sensing sensitive targets in vehicle-side images. Furthermore, this application utilizes a pre-set weight adjustment strategy to dynamically adjust the weights of each module based on the detection state, trajectory stability, motion continuity, and prediction error of the current image frame. Simultaneously, this application performs weighted fusion of the detection boxes corresponding to each module with the weights, and uses the weighted fused detection boxes as the final detection result. Thus, this application can achieve accurate perception and stable localization of sensitive targets in different vehicle-side scenarios. Therefore, this application can effectively suppress detection box jitter, reduce the target false negative rate, and improve localization stability. This solves the technical problems of insufficient robustness and unstable detection and localization in existing target detection methods for complex vehicle-side environments.

[0129] Example 3

[0130] Figure 7 A vehicle-mounted sensing sensitive target desensitization device according to this embodiment is shown, which corresponds to the method described according to Embodiment 1. Reference Figure 7 As shown, the device includes: a processor 710; and a memory 720 connected to the processor 710, used to provide the processor 710 with instructions to perform the following processing steps: inputting the acquired vehicle-end image into a pre-built information perception model for sensitive target perception, wherein the information perception model includes a target detection network for detecting sensitive targets, a target tracking module for recording the trajectory of sensitive targets, an optical flow motion estimation module for estimating the cross-frame motion trend of sensitive targets, and a state prediction module for predicting the future motion state of sensitive targets, and the vehicle-end image includes multiple image frames; determining a first detection box corresponding to the sensitive target in each image frame based on the target detection network, determining a second detection box corresponding to the sensitive target in each image frame based on the target tracking module, and determining a second detection box corresponding to the sensitive target in each image frame based on the optical flow motion estimation module; and determining a second detection box corresponding to the sensitive target in each image frame based on the optical flow motion estimation module. The flow motion estimation module determines the third detection box corresponding to the sensitive target within each image frame, and the state prediction module determines the fourth detection box corresponding to the sensitive target within each image frame. Based on a pre-set weight adjustment strategy, the first weight corresponding to the first detection box, the second weight corresponding to the second detection box, the third weight corresponding to the third detection box, and the fourth weight corresponding to the fourth detection box are determined respectively. According to the first weight, the second weight, the third weight, and the fourth weight, the first detection box, the second detection box, the third detection box, and the fourth detection box are weighted and fused to output the final detection box corresponding to the sensitive target within each image frame. The sensitive target region in the vehicle image is located according to the final detection box, and the sensitive target region is desensitized according to a preset desensitization method.

[0131] Optionally, the weight adjustment strategy includes: associating a first weight with the confidence score output by the target detection network; associating a second weight with the duration of the motion trajectory of the sensitive target; associating a third weight with the motion consistency of the sensitive target; and associating a fourth weight with the prediction error generated by the state prediction module.

[0132] Optionally, the weight adjustment strategy further includes: classifying the current vehicle scene into static scenes, high-motion scenes, nighttime scenes, and blurred scenes based on the environmental features included in the vehicle-end image; in static scenes, increasing the first weight of the target detection network and decreasing the third weight of the optical flow motion estimation module; in high-motion scenes, increasing the second weight of the target tracking module and the third weight of the optical flow motion estimation module; in nighttime scenes, increasing the second weight of the target tracking module and the fourth weight of the state prediction module; and in blurred scenes, decreasing the first weight of the target detection network and increasing the second weight of the target tracking module and the fourth weight of the state prediction module.

[0133] Optionally, the environmental features include brightness features, sharpness features, global motion features, and sensitive target motion features. The operation of classifying the current vehicle scene into static scenes, high-motion scenes, nighttime scenes, and blurred scenes based on the environmental features included in the vehicle-end image includes: extracting the brightness features, sharpness features, global motion features, and sensitive target motion features of the current image frame; determining the current vehicle scene as a nighttime scene when the ambient brightness corresponding to the brightness feature is lower than a first preset threshold; determining the current vehicle scene as a blurred scene when the image sharpness corresponding to the sharpness feature is lower than a second preset threshold; determining the current vehicle scene as a high-motion scene when the inter-frame motion intensity corresponding to the global motion feature is higher than a third preset threshold; and determining the current vehicle scene as a static scene when the inter-frame motion intensity is lower than the third preset threshold and the target motion intensity corresponding to the sensitive target motion feature is lower than a fourth preset threshold.

[0134] Optionally, when the detection of a sensitive target fails, the information perception model uses the third detection box, the first recovery weight coefficient corresponding to the third detection box, the fourth detection box, and the second recovery weight coefficient corresponding to the fourth detection box to perform secondary fusion on the current image frame of the vehicle-end image, and uses the fused detection box as the final detection box of the sensitive target in the current image frame.

[0135] Optionally, the operation of locating the sensitive target region in the vehicle-end image based on the final detection box and performing desensitization processing on the sensitive target region based on a preset desensitization method includes: determining the image region enclosed by the final detection box within the current image frame of the vehicle-end image based on the coordinate information of the final detection box, and marking the image region as a sensitive target region; and performing desensitization operation on the sensitive target region according to the preset desensitization method, while keeping the original image data of the non-sensitive target regions within the current image frame unchanged.

[0136] Optionally, the sensitive target area is desensitized according to a preset desensitization method, including: receiving a string key input by the user, generating a fixed-length hash value using a hash algorithm, and mapping the hash result to encryption control parameters, including the initial value and control parameters of the SLCS chaotic system and diffusion and scrambling control parameters; generating a key stream using the SLCS chaotic system based on the initial value and control parameters; constructing a cryptographic matrix based on the key stream and parameterizing the cryptographic matrix according to the diffusion and scrambling control parameters; performing a forward diffusion operation on the sensitive target area to couple pixel values ​​with neighboring pixels and the cryptographic matrix; generating a dynamic scrambling index using the cryptographic matrix and performing a non-linear scrambling operation on the diffused pixel sequence; and performing a backward diffusion operation on the scrambling result to convert the sensitive target area into an unrecognizable ciphertext image area and replace the corresponding area in the original vehicle-end image.

[0137] Therefore, according to this embodiment, this application introduces a target tracking module, an optical flow motion estimation module, and a state prediction module on the basis of the target detection network to construct an information perception model for sensing sensitive targets in vehicle-side images. Furthermore, this application utilizes a pre-set weight adjustment strategy to dynamically adjust the weights of each module based on the detection state, trajectory stability, motion continuity, and prediction error of the current image frame. Simultaneously, this application performs weighted fusion of the detection boxes corresponding to each module with the weights, and uses the weighted fused detection boxes as the final detection result. Thus, this application can achieve accurate perception and stable localization of sensitive targets in different vehicle-side scenarios. Therefore, this application can effectively suppress detection box jitter, reduce the target false negative rate, and improve localization stability. This solves the technical problems of insufficient robustness and unstable detection and localization in existing target detection methods for complex vehicle-side environments.

[0138] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0139] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0140] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0141] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0142] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0143] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0144] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for desensitizing vehicle-mounted sensing sensitive targets, characterized in that, include: The acquired vehicle-side image is input into a pre-built information perception model for sensitive target perception. The information perception model includes a target detection network for detecting sensitive targets, a target tracking module for recording the trajectory of the sensitive targets, an optical flow motion estimation module for estimating the cross-frame motion trend of the sensitive targets, and a state prediction module for predicting the future motion state of the sensitive targets. The vehicle-side image includes multiple image frames. Based on the target detection network, a first detection box corresponding to the sensitive target in each image frame is determined; based on the target tracking module, a second detection box corresponding to the sensitive target in each image frame is determined; based on the optical flow motion estimation module, a third detection box corresponding to the sensitive target in each image frame is determined; and based on the state prediction module, a fourth detection box corresponding to the sensitive target in each image frame is determined. Based on a pre-set weight adjustment strategy, a first weight corresponding to the first detection box, a second weight corresponding to the second detection box, a third weight corresponding to the third detection box, and a fourth weight corresponding to the fourth detection box are determined respectively. Based on the first weight, the second weight, the third weight, and the fourth weight, the first detection box, the second detection box, the third detection box, and the fourth detection box are weighted and fused to output the final detection box corresponding to the sensitive target in each image frame; as well as The sensitive target region in the vehicle image is located based on the final detection box, and the sensitive target region is desensitized based on a preset desensitization method.

2. The method according to claim 1, characterized in that, The weight adjustment strategy includes: The first weight is correlated with the confidence score output by the target detection network; The second weight is associated with the duration of the motion trajectory of the sensitive target; Associate the third weight with the motion consistency of the sensitive target; and The fourth weight is associated with the prediction error generated by the state prediction module.

3. The method according to claim 1, characterized in that, The weight adjustment strategy also includes: Based on the environmental features included in the vehicle-side image, the current vehicle-side scene is divided into static scene, high-motion scene, night scene, and blurred scene; In the static scene, the first weight of the target detection network is increased, and the third weight of the optical flow motion estimation module is decreased; In the high-motion scenario, the second weight of the target tracking module and the third weight of the optical flow motion estimation module are increased; In the nighttime scene, the second weight of the target tracking module and the fourth weight of the state prediction module are increased; and In the blurred scene, the first weight of the target detection network is reduced, and the second weight of the target tracking module and the fourth weight of the state prediction module are increased.

4. The method according to claim 3, characterized in that, The environmental features mentioned include brightness features, sharpness features, global motion features, and sensitive target motion features. The operation of classifying the current vehicle scene into static scenes, high-motion scenes, nighttime scenes, and blurred scenes based on the environmental features included in the vehicle image includes: Extract the brightness features, sharpness features, global motion features, and sensitive target motion features of the current image frame; When the ambient brightness corresponding to the brightness feature is lower than the first preset threshold, the current vehicle scene is determined to be the night scene. When the image clarity corresponding to the clarity feature is lower than the second preset threshold, the current vehicle scene is determined to be the blurry scene; When the inter-frame motion intensity corresponding to the global motion feature is higher than a third preset threshold, the current vehicle-side scene is determined to be a high-motion scene; and When the motion intensity between consecutive frames is lower than the third preset threshold and the motion intensity of the target corresponding to the sensitive target motion feature is lower than the fourth preset threshold, the current vehicle scene is determined to be the static scene.

5. The method according to claim 4, characterized in that, When the detection of the sensitive target fails, the information perception model uses the third detection box, the first recovery weight coefficient corresponding to the third detection box, the fourth detection box, and the second recovery weight coefficient corresponding to the fourth detection box to perform secondary fusion on the current image frame of the vehicle image, and uses the fused detection box as the final detection box of the sensitive target in the current image frame.

6. The method according to claim 2, characterized in that, The operation of locating the sensitive target region in the vehicle image based on the final detection box and performing desensitization processing on the sensitive target region based on a preset desensitization method includes: Based on the coordinate information of the final detection box, the image region enclosed by the final detection box is determined within the current image frame of the vehicle-end image, and the image region is marked as the sensitive target region; and The sensitive target area is desensitized according to the preset desensitization method, while the original image data of the non-sensitive target area in the current image frame remains unchanged.

7. The method according to claim 6, characterized in that, The sensitive target area is desensitized according to a preset desensitization method, including: The system receives a string key input by the user, generates a fixed-length hash value using a hash algorithm, and maps the hash result to encryption control parameters, which include the initial value and control parameters of the SLCS chaotic system, as well as diffusion and scrambling control parameters. Based on the initial value and the control parameters, a key stream is generated using the SLCS chaotic system; A cryptographic matrix is ​​constructed based on the key stream, and the cryptographic matrix is ​​parameterized according to the diffusion and scrambling control parameters; A forward diffusion operation is performed on the sensitive target region to couple the pixel value with the neighboring pixels and the cryptographic matrix; The cryptographic matrix is ​​used to generate a dynamic scrambling index, which is then used to perform a non-linear scrambling operation on the diffused pixel sequence; and The scrambling result is back-diffused to convert the sensitive target region into an unrecognizable encrypted image region and replace the corresponding region in the original vehicle image.

8. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, the method described in any one of claims 1 to 7 is performed by a processor.

9. A vehicle-mounted sensing sensitive target desensitization device, characterized in that, include: The sensitive target perception module is used to input the acquired vehicle-side image into a pre-built information perception model for sensitive target perception. The information perception model includes a target detection network for detecting sensitive targets, a target tracking module for recording the trajectory of the sensitive targets, an optical flow motion estimation module for estimating the cross-frame motion trend of the sensitive targets, and a state prediction module for predicting the future motion state of the sensitive targets. The vehicle-side image includes multiple image frames. The detection box determination module is used to determine a first detection box corresponding to the sensitive target in each image frame based on the target detection network, a second detection box corresponding to the sensitive target in each image frame based on the target tracking module, a third detection box corresponding to the sensitive target in each image frame based on the optical flow motion estimation module, and a fourth detection box corresponding to the sensitive target in each image frame based on the state prediction module. The weight determination module is used to determine, based on a pre-set weight adjustment strategy, a first weight corresponding to the first detection box, a second weight corresponding to the second detection box, a third weight corresponding to the third detection box, and a fourth weight corresponding to the fourth detection box. The final detection box output module is used to perform weighted fusion of the first detection box, the second detection box, the third detection box, and the fourth detection box according to the first weight, the second weight, the third weight, and the fourth weight, and output the final detection box corresponding to the sensitive target in each image frame; as well as The desensitization module is used to locate the sensitive target area in the vehicle image based on the final detection box, and to perform desensitization processing on the sensitive target area based on a preset desensitization method.

10. A vehicle-mounted sensing sensitive target desensitization device, characterized in that, include: processor; as well as A memory, connected to the processor, for providing the processor with instructions to perform the following processing steps: The acquired vehicle-side image is input into a pre-built information perception model for sensitive target perception. The information perception model includes a target detection network for detecting sensitive targets, a target tracking module for recording the trajectory of the sensitive targets, an optical flow motion estimation module for estimating the cross-frame motion trend of the sensitive targets, and a state prediction module for predicting the future motion state of the sensitive targets. The vehicle-side image includes multiple image frames. Based on the target detection network, a first detection box corresponding to the sensitive target in each image frame is determined; based on the target tracking module, a second detection box corresponding to the sensitive target in each image frame is determined; based on the optical flow motion estimation module, a third detection box corresponding to the sensitive target in each image frame is determined; and based on the state prediction module, a fourth detection box corresponding to the sensitive target in each image frame is determined. Based on a pre-set weight adjustment strategy, a first weight corresponding to the first detection box, a second weight corresponding to the second detection box, a third weight corresponding to the third detection box, and a fourth weight corresponding to the fourth detection box are determined respectively. Based on the first weight, the second weight, the third weight, and the fourth weight, the first detection box, the second detection box, the third detection box, and the fourth detection box are weighted and fused to output the final detection box corresponding to the sensitive target in each image frame; as well as The sensitive target region in the vehicle image is located based on the final detection box, and the sensitive target region is desensitized based on a preset desensitization method.

Images