Camera smudginess detection method and system and storage medium

Abstract

The invention provides a camera smudginess detection method and system and a storage medium, and the method comprises the steps: obtaining image data, and inputting the image data to a smudginess detection model; acquiring a feature image from the image data through the smudginess detection model, identifying a target square frame where smudginess is located based on the feature image, and outputting the position and confidence score of the target square frame; according to the position of the target square frame and the confidence score, calculating to obtain a prediction probability of dirt shielding; and obtaining a camera smudginess detection result according to the prediction probability. According to the method provided by the invention, the accuracy of camera smudginess shielding detection is effectively improved, and the stability and reliability of smudginess detection result output are effectively improved.

Description

Technical Field

[0001] The present application belongs to the field of computer vision technology, and specifically relates to a camera dirt detection method, system and storage medium. Background Art

[0002] With the widespread use of cameras in various applications, image quality has become a key factor. In many autonomous driving or assisted driving functions, it is necessary to rely on images captured by the camera as input, and then analyze and process these images. Camera dirtiness will seriously affect the perception performance of the perception system, causing the captured images to be inconsistent with the actual situation around the car. Unreliable image information may cause the autonomous driving or assisted driving functions to make wrong judgments, which in turn leads to safety issues. Therefore, it is particularly important to effectively detect camera dirtiness. Camera dirtiness detection requires detecting whether the camera is dirty and sending the detection results to the perception module that needs to use the camera data to determine whether to continue its perception function. Common dirtiness detection is to classify the images of the detected camera into dirty pictures or clear pictures. However, this classification is a global judgment of the entire picture. When the object to be detected is not obvious in the image and the avoidance area accounts for a small proportion, this classification will result in misjudgment. In addition, this type of classification algorithm has no result smoothing function, resulting in unstable classification output results. Summary of the invention

[0003] In order to solve the above technical problems, the present application proposes a camera dirt detection method, system and storage medium with high dirt detection accuracy and stable prediction result output.

[0004] Specifically, the present application proposes a camera dirt detection method, comprising: Acquire image data and input the image data into a dirt detection model. Acquire a feature image from the image data through the dirt detection model, identify a target box where dirt is located based on the feature image, and output the position and confidence score of the target box. Calculate the predicted probability of dirt occlusion based on the position and confidence score of the target box; and obtain a camera dirt detection result based on the predicted probability.

[0005] In the above technical solution, the dirt detection model identifies the target box where the dirt is located from the feature image and outputs the position and confidence score of the target box, which effectively improves the accuracy of the camera dirt detection. The predicted probability is obtained by calculating the position and confidence score of the target box, and the final camera dirt occlusion detection result is obtained according to the predicted probability, which effectively improves the stability and reliability of the camera dirt occlusion detection result.

[0006] As an implementation manner, the dirt detection model is at least a convolutional neural network, and acquiring the feature image includes: According to the convolutional neural network, a variety of feature information is obtained from the image data. The multiple feature information is fused to obtain the feature image. The feature information includes at least image features and semantic features. The convolutional neural network is used to obtain a variety of feature information from the image data, so that a more in-depth analysis of complex images is performed. By fusing a variety of feature information, a variety of feature information of the image is comprehensively considered, which effectively improves the accuracy of dirt occlusion detection and enhances the ability to identify dirt.

[0007] Furthermore, the identifying the target box where the dirt is located based on the feature image includes: The feature image is divided into a preset number of grids, and the target box where the dirt is located is identified according to the feature image after grid division, and the position and confidence score of the target box are obtained; wherein, obtaining the confidence score includes: obtaining the probability value of the target box according to the convolutional neural network, and calculating the probability value according to a preset function to obtain the confidence score.

[0008] By dividing the feature image into a preset number of grids, the image can be analyzed more carefully, so that the location of the dirt can be more accurately located in the local area, thereby improving the accuracy of detection. By calculating the probability value, the reliability of target box recognition is improved.

[0009] Furthermore, obtaining the predicted probability of dirt occlusion includes: The target box is screened according to a preset confidence threshold. The screened target box is tracked in real time through a target motion model. The occlusion influence degree is calculated according to the confidence score and the target box tracked in real time. The predicted probability of the dirt occlusion is obtained according to the occlusion influence degree.

[0010] Among them, the accuracy of the prediction result is ensured by screening the target box, and the change of the dirtiness state can be captured in time by real-time tracking the screened target box.

[0011] Furthermore, the real-time tracking of the filtered target box by using the target motion model includes: The first position and area of ​​the current target box are recorded in real time.

[0012] The second position of the target box in the next frame is predicted by the target motion model.

[0013] The overlapping area ratio between the current target box and the target box of the next frame is calculated according to the first position, the second position and the area.

[0014] Determine whether the overlapping area ratio is greater than a preset threshold. If so, determine that the target box is tracked successfully and retain the target box of the next frame; otherwise, determine that the target box tracking fails, remove the target box of the next frame and re-predict the position of the target box of the next frame according to the target motion model until the tracking is successful.

[0015] Among them, by calculating the overlap area ratio between the current target box and the target box of the next frame, it is possible to effectively determine whether the target box of the next frame is within the field of view, thereby improving the accuracy of target box tracking. By eliminating the target box that failed to be tracked and re-predicting, the tracking strategy can be adaptively adjusted, improving the robustness of camera dirt detection. And by real-time tracking of the target box, the real-time performance of camera dirt detection is improved.

[0016] Furthermore, the calculation of the occlusion influence degree includes: The area ratio of the current target box area to the entire image is obtained. The product of the confidence score and the area ratio is calculated, and the product is used as the occlusion influence degree.

[0017] Furthermore, obtaining the predicted probability of dirt occlusion according to the occlusion influence degree includes: The average value of the occlusion influence of all target boxes in the current frame is calculated, and the average value is used as the predicted probability of the dirt occlusion. By calculating the average value of the occlusion influence of all target boxes, the occlusion situation of the entire camera can be comprehensively evaluated, which reduces the misjudgment caused by the dirt occlusion of a single target and improves the accuracy and stability of the camera's dirt occlusion detection.

[0018] Further, obtaining the camera dirtiness detection result according to the predicted probability includes: The predicted probability change value is obtained according to the predicted probability of the current frame and the predicted probability of the previous frame. When the predicted probability change value is less than the preset change threshold, and the predicted probability of the current frame is greater than the preset probability threshold, it is determined that the current camera is blocked by dirt. The stability of the dirt detection result is ensured by obtaining the predicted probability change value and judging the predicted probability change value.

[0019] Based on the same inventive concept, the present application also proposes a system for detecting camera dirtiness, the system comprising: The data acquisition module is used to acquire image data and input the image data into the dirt detection model.

[0020] The detection module is used to obtain a feature image from the image data through the dirt detection model, identify a target box where dirt is located based on the feature image, and output a position and a confidence score of the target box.

[0021] The calculation module is used to calculate the predicted probability of dirt occlusion according to the position and confidence score of the target box.

[0022] And, a judgment module is used to obtain a camera dirtiness detection result according to the predicted probability.

[0023] Based on the same inventive concept, the present application also proposes a computer-readable storage medium, which stores computer-executable instructions. The computer-executable instructions can be read by a domain controller and execute the camera dirtiness detection method.

[0024] Compared with the prior art, this application has at least the following beneficial effects: This application uses a dirt detection model to extract feature images from the acquired image data, and identifies the location and confidence score of the target box where the dirt is located based on the feature image, effectively improving the accuracy of camera dirt detection. By calculating the predicted probability of dirt occlusion and obtaining the camera dirt detection result based on the predicted probability, the camera dirt detection result is made more reliable, effectively improving the stability and reliability of the camera dirt detection result output. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of a camera dirt detection method shown in an embodiment of the present application.

[0026] Figure 2 It is a schematic diagram of a camera dirt detection system shown in an embodiment of the present application. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0028] Embodiment 1: Please refer to Figure 1 , the camera dirt detection method mainly includes: Acquire image data and input the image data into a dirt detection model. The dirt detection model may be at least a convolutional neural network (CNN). The convolutional neural network may be provided with an image encoder, which is mainly composed of a convolutional layer, an activation layer, and a pooling layer of the convolutional neural network. The image encoder extracts feature information of the input image data, such as edge features, texture features, and shape features, through a convolution operation.

[0029] The dirt detection model obtains a feature image from the image data, identifies the target box where the dirt is located based on the feature image, and outputs the position and confidence score of the target box. The feature image can be mainly obtained by fusing multiple feature information. For example, the dirt detection model can extract image features and semantic features from the input image data, and the image features and semantic features are superimposed and fused to obtain the feature image. The target box can be mainly divided into N grids by the feature image, and the grid where the dirt is located is identified according to the dirt detection model, and the target box is obtained according to the grid where the dirt is located, where N is a natural number. The confidence score is mainly the confidence score of each detection category obtained by normalizing the unnormalized probability value output by the dirt detection model through the softmax activation function. Each detection category mainly includes the camera background part and the dirt occlusion part. The softmax activation function is a key activation function in the multi-class classification problem. The output data of the dirt detection model can be converted into a probability distribution through the softmax activation function, and the probability distribution is normalized to obtain the confidence score.

[0030] The predicted probability of dirt occlusion is calculated based on the position and confidence score of the target box. The target box is tracked in real time based on its position, the ratio of the area of ​​the camera image occupied by the target box is calculated, and then the ratio is multiplied by the confidence score to obtain the degree of occlusion of the target box on the picture. The predicted probability of dirt occlusion is obtained by taking the average of all occlusion influences of the current frame.

[0031] The camera dirtiness detection result is obtained according to the predicted probability. According to the comparison between the predicted probability of the current frame being blocked by dirt and the predicted probability of the previous frame being blocked by dirt, the predicted probability change value of dirtiness is obtained, and it is judged whether the probability change value is less than the preset change threshold and whether the current frame predicted probability is greater than the preset probability threshold. If so, the obtained dirtiness detection result is that the camera is currently blocked by dirt, otherwise, the camera is not currently blocked. Among them, those skilled in the art can set the preset change threshold to 10% and the preset probability threshold to 70% according to actual needs. For example, when the predicted probability of the current frame is 71% and the probability change value is 9%, the probability change value of the current frame is less than the probability change threshold, the predicted probability is greater than the preset probability threshold, and the camera is currently blocked by dirt.

[0032] In some embodiments, the dirt detection model is at least a convolutional neural network, and acquiring the feature image includes: A plurality of feature information is obtained from the image data according to the convolutional neural network, and the plurality of feature information is fused to obtain the feature image, wherein the feature information includes at least image features and semantic features.

[0033] The convolutional neural network may also include an image encoder, which performs a convolution operation on the image data to obtain multiple feature information in the image data. The image features may include at least image position and image texture features. The semantic features are higher-level semantic information captured mainly by a deeper network structure, such as object category, position and contextual relationship. The fusion of multiple feature information is mainly to generate the feature image after superimposing and fusing different feature information in the channel dimension.

[0034] In some embodiments, the step of identifying the target box where the dirt is located based on the feature image comprises: The feature image is divided into a preset number of grids, and the target box where the dirt is located is identified according to the feature image after the grids are divided, and the position and confidence score of the target box are obtained. The obtaining of the confidence score includes: obtaining the probability value of the target box according to the convolutional neural network, and calculating the probability value according to a preset function to obtain the confidence score.

[0035] Among them, those skilled in the art can set the preset number to 16x16 or 8x8 according to actual needs, but it is not limited to this. For example, the feature image is divided into 8x8 grids. Dirt recognition is performed on each divided grid, and after the dirt is recognized, the grid where the dirt is located is returned, and the size and position of the target box are obtained according to the grid where the dirt is located, and the target box is obtained according to the size and position of the target box. The dirt can mainly be silt and stains. The dirt probability value of each target box is obtained through the output of the convolutional neural network. The preset function can mainly be a softmax activation function, and the confidence score is obtained by calculating the softmax activation function. Through the above implementation, the dirt detection model can efficiently and accurately locate the dirt in the image and obtain the confidence score, so that in practical applications, such as industrial inspection, environmental monitoring and other scenarios, it can quickly judge and deal with potential camera dirt occlusion problems.

[0036] In some embodiments, obtaining the predicted probability of dirt occlusion includes: The target box is screened according to a preset confidence threshold, the screened target box is tracked in real time through a target motion model, the occlusion influence degree is calculated according to the confidence score and the target box tracked in real time, and the predicted probability of the dirty occlusion is obtained according to the occlusion influence degree.

[0037] Among them, those skilled in the art can set the confidence threshold to 40% according to actual needs, but it is not limited thereto. According to the confidence threshold, all target boxes in the image of the current frame are screened, and the target boxes whose confidence is greater than the confidence threshold are retained, and all target boxes whose confidence is less than or equal to the confidence threshold are eliminated.

[0038] In some embodiments, the real-time tracking of the filtered target box by using the target motion model includes: The first position and area of ​​the current target box are recorded in real time, and the second position of the target box of the next frame is predicted by the target motion model. The overlapping area ratio of the current target box and the target box of the next frame is calculated according to the first position, the second position and the area.

[0039] Determine whether the overlapping area ratio is greater than a preset threshold. If so, determine that the target box is tracked successfully and retain the target box of the next frame; otherwise, determine that the target box tracking fails, remove the target box of the next frame and re-predict the position of the target box of the next frame according to the target motion model until the tracking is successful.

[0040] Wherein, the target motion model can mainly be a uniform target motion model in a Kalman filter, which is a basic model for dynamic system evaluation and is suitable for tracking targets in motion. The uniform target motion model tracks the position, area and aspect ratio of the target box by assuming that the target moves at a constant speed in a short time, and predicts the position of the target box in the next frame image through a state transfer matrix, an observation matrix and a covariance matrix. Wherein, the state transfer matrix describes how the target box is transferred from a current frame to the next frame, and dynamically records the position of the target box. The observation matrix maps the state space to the observation space, indicating the amount that can be observed from the state, for example, the position and area of ​​the target box of the next frame can be predicted through the observation matrix. The covariance matrix is ​​mainly used to describe the uncertainty of the state and observation. For example, the noise and uncertainty in the observation, the uncertainty of the prediction of the target box of the next frame. The uniform target motion model can dynamically adjust the position estimation of the target by prediction and update in the case of state uncertainty, thereby improving the accuracy of target box tracking.

[0041] The first position, second position and area of ​​the target box are used to calculate the overlap area ratio between the current target box and the target box of the next frame, and whether the tracking is successful is determined based on the overlap area. For example, if the first position of the target box of the current frame is (2, 2), the second position of the target box of the next frame is predicted to be (4, 4) by the target motion model, and the area of ​​the target box is 16, the overlap area ratio is calculated to be 25%, and the preset threshold is 20%. Therefore, the overlap area ratio is greater than the preset threshold, so the target box is tracked successfully, and the target box of the next frame predicted by the target motion model is retained.

[0042] In some embodiments, the calculating the degree of occlusion influence includes: The area ratio of the current target box area to the overall image is obtained, and the product of the confidence score and the area ratio is calculated, so as to use the product as the degree of occlusion. For example, the area ratio of the current target box area to the overall camera image is 10%, and the current confidence score is 0.8, so the degree of occlusion of the current target box can be 8%.

[0043] In some embodiments, obtaining the predicted probability of dirt occlusion according to the occlusion influence degree includes: Calculate the average of the occlusion influence of all target boxes in the current frame, and use the average as the predicted probability of the dirty occlusion. For example, assuming that there are 3 target boxes in the current frame, and the occlusion influence of each target box is 80%, 90%, and 60% respectively, the average value is 77% to obtain the predicted probability.

[0044] In some embodiments, obtaining the camera dirtiness detection result according to the predicted probability includes: The predicted probability change value is obtained according to the predicted probability of the current frame and the predicted probability of the previous frame. When the predicted probability change value is less than a preset change threshold and the predicted probability of the current frame is greater than a preset probability threshold, it is determined that the current camera is blocked by dirt.

[0045] Among them, those skilled in the art can set the preset change threshold to 10% and the preset probability threshold to 70% according to actual needs, but are not limited thereto. For example, if the prediction probability of the current frame is 77% and the prediction probability of the previous frame is 75%, then the prediction probability of the current frame is compared with the prediction probability of the previous frame, and the predicted probability change value obtained is 2%. The predicted probability and the predicted probability change value are compared according to the preset change threshold and the preset probability threshold, and the current camera dirt detection result is obtained as the camera is dirty and blocked.

[0046] For example, if the prediction probability of the current frame is 68% and the prediction probability of the previous frame is 60%, the prediction probability of the current frame is compared with the prediction probability of the previous frame, and the prediction probability change value is 8%. The prediction probability and the prediction probability change value are compared according to the preset change threshold and the preset probability threshold, and the prediction probability of the current frame is less than the preset probability threshold, then there is no dirt occlusion in the current camera.

[0047] In the specific implementation process, taking the dirt occlusion detection of the front camera of the vehicle as an example, the image data is obtained through the front camera of the vehicle, and the image data is input into the convolutional neural network. An image encoder is set in the convolutional neural network. The image encoder extracts multiple features in the image data and superimposes and fuses the multiple features into a feature image. The target box where the dirt is located is identified based on the feature image, and the position and confidence score of the target box are obtained. The target box is tracked in real time in the image data through the target motion model, and the degree of dirt occlusion influence is calculated by the confidence score to be 78%. The average value of all occlusion influences in the current frame is 71%, and the predicted probability of dirt occlusion in the current frame is 71%. The predicted probability of dirt occlusion in the previous frame is detected to be 76%. The predicted probability and probability change value of the current frame are compared according to the preset probability threshold of 70% and the preset change threshold of 10%, and the dirt detection result obtained is that there is dirt occlusion in the front camera of the vehicle.

[0048] Embodiment 2: Please refer to Figure 2 The present application also proposes a camera dirt detection system, which mainly includes: a data acquisition module, a detection module, a calculation module and a judgment module.

[0049] The data acquisition module is used to acquire image data and input the image data into the dirt detection model. The image data can be each frame of image data in the video screen taken by the camera, and the dirt detection model is mainly a convolutional neural network.

[0050] The detection module is used to obtain a feature image from the image data through the dirt detection model, identify the target box where the dirt is located based on the feature image, and output the position and confidence score of the target box. In the dirt detection model, the image data is mainly convolved by an image encoder to obtain a variety of different types of feature information, such as image features and semantic features. Among them, image features can mainly be image texture features, position features and shape features. Then, the various types of features are overlapped and fused to obtain the feature image. The target box is mainly obtained by dividing the feature image in the dirt detection model into N grids, and obtaining the target box where the dirt is located according to the feature image after the grid is divided.

[0051] The calculation module is used to calculate the predicted probability of dirty occlusion according to the position and confidence score of the target box. The position of the target box of the next frame is predicted by the target motion model based on the position of the target box, and the overlapping area ratio of the target box of the next frame and the current target box is calculated to track the target box. The occlusion influence degree of the target box is calculated based on the confidence score, and the average occlusion influence degree of all target boxes in the current frame is calculated to obtain the predicted probability of dirty occlusion.

[0052] And, the judgment module is used to obtain the camera dirt detection result according to the predicted probability. The predicted probability change value is obtained mainly through the predicted probability of dirt occlusion in the current frame and the predicted probability of dirt occlusion in the previous frame, and the predicted probability and the predicted probability change value are judged according to the preset probability threshold and the preset probability change threshold to obtain the camera dirt detection result.

[0053] Embodiment three: The present application also proposes a computer-readable storage medium, the computer-readable storage medium comprising: The computer-readable storage medium stores computer-executable instructions.

[0054] The computer executable instructions implement the camera dirt detection method when executed by the control processor.

[0055] In the computer-readable storage medium, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk SolidState Disk (SSD)), etc.

[0056] In summary, the present application uses a dirt detection model to extract feature images from image data, and identifies the location and confidence score of the target box where the dirt is located based on the feature image, thereby solving the problem of misjudgment of dirt in camera dirt detection and effectively improving the accuracy of camera dirt detection. By calculating the predicted probability of dirt occlusion and obtaining the camera dirt detection result based on the predicted probability, the problem of unstable camera dirt detection results is solved, effectively improving the stability and reliability of the output of the camera dirt detection result.

[0057] In several embodiments provided in the present application, it is understood that each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved.

[0058] If the function is implemented in the form of a software function module 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 application can essentially or in other words, the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for an electronic device to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

[0059] The specific embodiments described above further describe the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the scope of protection of the present application. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A camera dirt detection method, characterized in that: include: Acquiring image data, and inputting the image data into a dirt detection model; Acquire a feature image from the image data through the dirt detection model, identify a target box where dirt is located based on the feature image, and output a position and a confidence score of the target box; Calculating the predicted probability of dirt occlusion according to the position and confidence score of the target box; and A camera dirtiness detection result is obtained according to the predicted probability.

2. The camera dirt detection method according to claim 1, characterized in that: The dirt detection model is at least a convolutional neural network, and acquiring a feature image includes: Acquiring a variety of feature information from the image data according to the convolutional neural network; Fusing the multiple feature information to obtain the feature image; The feature information includes at least image features and semantic features.

3. The camera dirt detection method according to claim 2, characterized in that: The step of identifying the target box where the dirt is located based on the feature image comprises: Dividing the feature image into a preset number of grids, identifying the target box where the dirt is located according to the feature image after the grid division, and obtaining the position and confidence score of the target box; Wherein, obtaining the confidence score includes: Obtaining a probability value of the target box according to the convolutional neural network; The probability value is calculated according to a preset function to obtain the confidence score.

4. The camera dirt detection method according to claim 1, characterized in that: The obtaining of the predicted probability of dirt occlusion includes: Screening the target box according to a preset confidence threshold; The screened target box is tracked in real time through the target motion model; Calculating the degree of occlusion impact according to the confidence score and the target box tracked in real time; The predicted probability of the dirt occlusion is obtained according to the occlusion influence degree.

5. The camera dirt detection method according to claim 4, characterized in that: The real-time tracking of the filtered target box by using the target motion model includes: Recording the first position and area of ​​the current target box in real time; Predicting a second position of the target box in the next frame by using the target motion model; Calculate the overlap ratio of the current target box and the target box of the next frame according to the first position, the second position and the area; Determine whether the overlapping area ratio is greater than a preset threshold. If so, determine that the target box is tracked successfully and retain the target box of the next frame; otherwise, determine that the target box tracking fails, remove the target box of the next frame and re-predict the position of the target box of the next frame according to the target motion model until the tracking is successful.

6. The camera dirt detection method according to claim 5, characterized in that: The calculating of the occlusion influence degree includes: Obtaining the area ratio of the target box to the entire image; The product of the confidence score and the area ratio is calculated, and the product is used as the occlusion influence degree.

7. The camera dirt detection method according to claim 4, characterized in that: The obtaining the predicted probability of dirt occlusion according to the occlusion influence degree includes: The average value of the occlusion influence degree of all target boxes in the current frame is calculated, and the average value is used as the predicted probability of the dirty occlusion.

8. The camera dirt detection method according to claim 1, characterized in that: The obtaining of the camera dirtiness detection result according to the predicted probability includes: Obtain a predicted probability change value based on the predicted probability of the current frame and the predicted probability of the previous frame; When the predicted probability change value is smaller than a preset change threshold, and the predicted probability of the current frame is greater than a preset probability threshold, it is determined that the current camera is blocked by dirt.

9. A system based on the camera dirt detection method according to any one of claims 1 to 8, characterized in that: The system comprises: A data acquisition module, used to acquire image data and input the image data into a dirt detection model; A detection module, configured to obtain a feature image from the image data by using the dirt detection model, identify a target box where dirt is located based on the feature image, and output a position and a confidence score of the target box; A calculation module, used for calculating the predicted probability of dirt occlusion according to the position and confidence score of the target box; And, a judgment module is used to obtain a camera dirtiness detection result according to the predicted probability.

10. A computer-readable storage medium storing computer-executable instructions, characterized in that: When the computer executable instructions are executed by the control processor, the camera dirt detection method as described in any one of claims 1-8 is implemented.

Images