A target detection method and device
By constructing a single-stage target detection network using the YOLOv5s network on a large commercial display screen, and combining the judgment of the phone's pose and stationary state, the problems of insufficient model size, speed and accuracy in the existing technology are solved, and fast and accurate mobile phone covert photography detection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING MYSHER TECH
- Filing Date
- 2022-08-01
- Publication Date
- 2026-06-19
AI Technical Summary
Existing anti-spy camera detection technologies suffer from insufficient model size, speed, and accuracy on large commercial display screens, making it difficult to meet the deployment requirements of the Android platform.
Using the YOLOv5s network as the base network, a single-stage object detection network is constructed through convolutional blocks, residual cascaded units, fast pooling units, and upsampling layers. Combined with feature map channel connection layers, the model is trained to identify the features of a hidden camera, including the phone's pose and stationary state, and uses IOU to determine whether the phone remains stationary.
It enables rapid and accurate detection of mobile phone surreptitious photography on large commercial display screens. The model is smaller and faster, suitable for deployment on the Android platform, and improves the accuracy and real-time performance of detection.
Smart Images
Figure CN115311742B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of computer vision, and particularly relates to the field of object detection, specifically a method for using object detection technology to prevent commercial display screens from being secretly photographed. Background Technology
[0002] With societal development and technological advancements, convenient and intelligent commercial display screens are gradually replacing traditional projectors as the primary video output device in conference rooms. In today's world, where the protection of business secrets is paramount, protecting the video information displayed on commercial display screens is equally important. Currently, however, there is a complete lack of solutions for preventing mobile phone filming on commercial display screens. This invention aims to propose a method for preventing mobile phone filming on commercial display screens, filling this gap in their application.
[0003] Existing anti-spy camera detection technologies are mainly based on two deep learning tasks: pose recognition and target detection.
[0004] Pose recognition identifies human postures in images, which can be used to determine if there is any attempt to prevent surreptitious filming. While pose recognition involves the detection of key human points, which is closer to the goal of preventing surreptitious filming, existing human pose recognition models can only be tested with a single person or a small group. In multi-person scenarios, the speed becomes extremely slow and the accuracy decreases.
[0005] Object detection is a method for detecting and identifying targets of interest in an image, enabling target localization and recognition. This method can be used to detect hidden cameras, and then additional criteria can be added to confirm whether a hidden camera is being used. However, existing anti-hidden camera target detection models are slow and large, making them extremely inconvenient for the Android platform ultimately intended for large commercial displays.
[0006] Meanwhile, existing computer vision-based anti-spy camera detection methods use relatively old network models, which are inferior to the newer networks in terms of model size, speed, and accuracy.
[0007] Therefore, there is an urgent need for a detection network with appropriate model size, high speed, and high accuracy, suitable for deployment on commercial display screens on the Android platform. Summary of the Invention
[0008] To address the shortcomings of existing technologies, the purpose of this application is to provide a method for preventing mobile phone-based unauthorized photography.
[0009] To achieve the above objectives, this application provides the following technical solution:
[0010] A target detection method includes the following steps:
[0011] S100: Acquire the image to be detected;
[0012] S200: Input the acquired image to be detected into the target detection network to determine whether the video image to be detected is in a preset specific state.
[0013] Preferably, the target detection network is a single-stage target detection network, which includes:
[0014] A network input layer, which is used to input the image to be detected;
[0015] A convolutional block, which consists of a convolutional layer, a batch normalization layer, and a sigmoid activation layer;
[0016] C3 block, wherein the C3 block is a residual cascaded unit composed of multiple convolutional blocks;
[0017] SPFF blocks, wherein the SPFF blocks are fast pooling units consisting of 2 convolutional layers and 3 max pooling layers;
[0018] An upsampling layer, wherein the upsampling layer uses the nearest neighbor method for upsampling;
[0019] Feature map channel connection layer;
[0020] The detection block consists of three convolutional activation units (CEUs). Each CEU processes output detection information at one scale level and comprises a convolutional layer and a sigmoid activation layer. After processing the detection information at three scale levels, multiple candidate boxes are obtained. Finally, the final target detection box is output using the maximum suppression method.
[0021] Preferably, the feature map channel connection layer is a concat layer.
[0022] Preferably, the single-stage target detection network is selected from the YOLO series networks.
[0023] Preferably, in step S200, the training steps of the anti-spy camera detection model include:
[0024] Step S200 further includes a model training step, specifically including:
[0025] S201: Collect mobile phone video images including those taken in a surreptitious state and those not taken in a surreptitious state, and divide them into training set and validation set;
[0026] S202: Input the training set into the anti-spy camera detection model to train the model. When the learning rate of the anti-spy camera detection model reaches the preset value, the training is complete.
[0027] Preferably, the model training step employs a pre-trained weight loading method to train the model, specifically including:
[0028] The dataset prepared in the data preparation steps is used as the network input. During model training, the initial learning rate lr0 is set to 0.01, and stochastic gradient descent (SGD) is used to update the network weights. The custom learning rate decay formula is as follows.
[0029] lr = lr0 ((1 – x / epochs) 0.99+0.01),
[0030] Where lr represents the learning rate for each iteration, x represents the current iteration number, and epochs represents the total number of iterations;
[0031] Preferably, the number of epochs is 300;
[0032] After the training is completed in step S202, a first detection network is obtained, which can initially eliminate non-surreptitious filming behavior.
[0033] Furthermore, in order to determine whether there was any surreptitious filming, the applicant summarized two important characteristics of surreptitious filming, especially surreptitious mobile phones: (1) the surreptitious mobile phone will be in a horizontal or vertical position to ensure that the angle is upright; (2) the surreptitious mobile phone will remain still to ensure that the focus is clear.
[0034] In view of the above-mentioned characteristics of surreptitious filming, preferably, step S200 further includes:
[0035] S203: Determination of voyeurism.
[0036] In another preferred embodiment, step S203 specifically includes the following steps:
[0037] S2031: Determine whether the mobile phone in the video image to be detected is horizontal or deviates from the horizontal by a certain angle, and whether it is vertical or deviates from the vertical by a certain angle.
[0038] In this step, after the model detects the phone, it obtains the phone's location and dimensions (x, y, w, h). x and y represent the top-left position of the phone's detection bounding box, and w and h represent the dimensions of the detection box. Since the detection box is close to the edge of the phone when it is placed horizontally or vertically, the ratio of the shorter side to the longer side of the detection box is minimized at this time. The ratio Rwh is as follows:
[0039] Rwh = min(w,h) / max(w,h)
[0040] When Rwh is less than the threshold T1, it can be preliminarily determined that the phone is in a state of covert filming; otherwise, the judgment ends. Regarding the value of T1, considering that most phones on the market have an aspect ratio of 16:9 or 16:10, meaning the ratio of the shorter side to the longer side of a typical phone is less than 0.625, the theoretical range for T1 is between 0.625 and 1.0. A T1 of 0.625 means that a small number of phones and some phones used for covert filming at small angles will be missed, resulting in a high false positive rate and a low false negative rate. A T1 of 1.0 means that most phones used for covert filming at small angles will be falsely detected, resulting in a low false positive rate and a high false negative rate. In balancing the false positive and false negative rates, the threshold T1 is chosen to be between 0.65 and 0.85. If T1 is less than 0.65, many phones will be missed, resulting in a high false negative rate. If T1 is greater than 0.85, many normally used phones will also be mistaken for covert filming phones, resulting in a high false negative rate.
[0041] Preferably, T1 can be selected as 0.75.
[0042] S2032: Determine whether the mobile phone in the video image to be detected remains stationary.
[0043] Preferably, step S2032 includes: determining the intersection-union ratio (IOU) of the detection boxes in the preceding and following frames, using the following formula:
[0044]
[0045] and
[0046]
[0047] Where (x1,y1,w1,h1) is the mobile phone detection box information of the previous frame; (x2,y2,w2,h2) is the mobile phone detection box information of the current frame, and the intersection area of the detection boxes is Iarea.
[0048] When the IOU is greater than the threshold T2, it indicates that the phone is actually stationary and there is a hidden camera incident; otherwise, the judgment ends. Furthermore, IOU=1.0 represents theoretical complete stillness, and IOU=0.0 indicates significant phone movement.
[0049] Preferably, considering factors such as hand tremors and model detection errors, it is difficult for the detection frames of two consecutive frames to achieve 100% overlap, i.e., complete stillness. Therefore, this embodiment sets an error range (0.03-0.08), that is, the IOU threshold T2 ranges from 0.92 to 0.97. If the value of T2 is less than 0.92, a normally used, slowly moving phone may be falsely detected as a hidden camera, resulting in a high false detection rate. If the value of T2 is greater than 0.97, a hidden camera may be missed due to external factors such as model, lighting, and slight hand tremors, resulting in a high missed detection rate.
[0050] Preferably, the value of T2 is 0.95, and actual test results show that this value range has the best anti-spy camera effect.
[0051] Furthermore, only when both features are simultaneously satisfied is it determined that mobile phone surreptitious filming has occurred. This invention determines the existence of surreptitious filming solely through preliminary S1 data processing and mobile phone detection information. Compared to other methods that utilize human body information, hand information, and mobile phone information to determine surreptitious filming, this model is smaller, faster, and better meets real-time requirements.
[0052] This application also provides a target detection device, comprising:
[0053] The acquisition unit is used to acquire the video image to be detected;
[0054] The judgment unit is used to input the acquired video image to be detected into the trained detection network and determine whether the target object in the video image to be detected is in a specific state according to the preset detection strategy.
[0055] Preferably, the target object is a mobile phone; the specific state is a covert filming state.
[0056] Preferably, the determining unit includes:
[0057] The model training subunit is used to train the anti-spy camera model;
[0058] The model validation subunit is used to validate the trained anti-spy camera model.
[0059] The input sub-unit is used to input the acquired video image to be detected into the trained anti-spy camera detection model;
[0060] The first judgment subunit is used to determine whether the mobile phone in the video image to be detected is horizontal or deviates from the horizontal by a certain angle and whether it is vertical or deviates from the vertical by a certain angle.
[0061] The second judgment subunit is used to determine whether the mobile phone in the video image to be detected remains stationary.
[0062] This application also provides a computer device, including:
[0063] Memory, used to store computer instructions;
[0064] A processor is used to execute computer instructions to implement the target detection method described above.
[0065] This application also provides an application of a method for preventing mobile phone spying, wherein the method is applied to a commercial display screen.
[0066] This application also provides an application of an anti-spy phone camera device, wherein the device is applied to a commercial display screen.
[0067] Compared with the prior art, the beneficial effects of this application are as follows:
[0068] By filtering the dataset and training and comparing the network, the detection accuracy of hidden cameras was improved. Finally, additional judgment criteria were added to definitively confirm whether hidden filming had occurred. Good detection results were achieved in actual testing.
[0069] Meanwhile, since the method proposed in this invention uses the YOLOv5s network, which has performed well in the past two years, as the base network for training the mobile phone detection model, the detection model trained by this network is smaller, faster, and more accurate, making it more suitable for deployment on the Android platform for commercial display screens, thus solving the problems existing in the prior art. Attached Figure Description
[0070] Figure 1 This is a flowchart of a method for preventing mobile phone spying, provided in one embodiment of this application;
[0071] Figure 2 This is a schematic diagram of the anti-spy camera detection model provided in another embodiment of this application;
[0072] Figure 3 This is a flowchart provided in another embodiment of the present application for determining whether a mobile phone in a video image to be detected is in a state of surreptitious filming;
[0073] Figure 4 This is an application flowchart of an anti-spy camera model provided in another embodiment of this application. Detailed Implementation
[0074] Specific embodiments of this application will now be described in detail with reference to the accompanying drawings. While specific embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.
[0075] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions in the specification are preferred embodiments for carrying out this application; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of this application. The scope of protection of this application shall be determined by the appended claims.
[0076] To facilitate understanding of the embodiments of this application, the following will provide further explanation and description with reference to the accompanying drawings and specific embodiments, and the accompanying drawings do not constitute a limitation on the embodiments of this application.
[0077] In one embodiment, such as Figure 1 As shown, a target detection method includes the following steps:
[0078] S100: Acquire the image to be detected;
[0079] S200: Input the acquired image to be detected into the target detection network to determine whether the video image to be detected is in a preset specific state.
[0080] Preferably, the preset specific state refers to the state of secretly filming.
[0081] In another embodiment, in step S200, the target detection network adopts a single-stage target detection network, which includes:
[0082] A network input layer, which is used to input the image to be detected;
[0083] A convolutional block, which consists of a convolutional layer, a batch normalization layer, and a sigmoid activation layer;
[0084] C3 block, wherein the C3 block is a residual cascaded unit composed of multiple convolutional blocks;
[0085] SPFF blocks, wherein the SPFF blocks are fast pooling units consisting of 2 convolutional layers and 3 max pooling layers;
[0086] An upsampling layer, wherein the upsampling layer uses the nearest neighbor method for upsampling;
[0087] Feature map channel connection layer;
[0088] The detection block consists of three convolutional activation units (CEUs). Each CEU processes output detection information at one scale level and comprises a convolutional layer and a sigmoid activation layer. After processing the detection information at three scale levels, multiple candidate boxes are obtained. Finally, the final target detection box is output using the maximum suppression method.
[0089] Preferably, the feature map channel connection layer is a concat layer.
[0090] Preferably, the single-stage target detection network is selected from the YOLO series networks; further, in the preferred embodiment, considering that the final network model will be deployed on the Android platform, the preferred embodiment only selects several lightweight networks from the YOLO series for comparison, as shown in Table 1.
[0091] Table 1 Comparison of the performance of YOLO series lightweight network models on the coco2017 validation set.
[0092]
[0093] As can be seen from the performance of several lightweight YOLO network models on the COCO2017 validation set shown in Table 1, the YOLOv5s network has the best detection performance. Therefore, this preferred embodiment selects YOLOv5s as the most preferred basic training network for this application.
[0094] The YOLOv5s network structure is as follows: Figure 2 As shown, the network input layer receives a 640 image. The network structure uses a 640-pixel RGB three-channel color image. The convolutional blocks consist of convolutional layers (conv2d), batch normalization layers (BatchNorm2d), and sigmoid activation layers (SiLU). The C3 block is a residual cascade unit composed of multiple convolutional blocks. The SPFF block is a spatial pyramid fast pooling unit consisting of two convolutional layers and three max-pooling layers. The upsampling layer uses the nearest neighbor method for upsampling. The concat layer is a feature map channel connection layer. The final detection block consists of three convolutional activation units, each processing one scale of output detection information. Each convolutional activation unit consists of one convolutional layer and one sigmoid activation layer. After processing the detection information at three scales, multiple candidate boxes are obtained. Finally, the final target detection box is output using the maximum suppression method.
[0095] In a further preferred embodiment, step S200 further includes a model training step, specifically including:
[0096] S201: Collect mobile phone video images including those taken in a surreptitious state and those not taken in a surreptitious state, and divide them into training set and validation set;
[0097] S202: Input the training set into the anti-spy camera detection model to train the model. When the learning rate of the anti-spy camera detection model reaches the preset value, the training is complete.
[0098] Preferably, the model training step employs a pre-trained weight loading method to train the model, specifically including:
[0099] The dataset prepared in the data preparation steps is used as the network input. During model training, the initial learning rate lr0 is set to 0.01, and stochastic gradient descent (SGD) is used to update the network weights. The custom learning rate decay formula is as follows.
[0100] lr = lr0 ((1 – x / epochs) 0.99+0.01),
[0101] Where lr represents the learning rate for each iteration, x represents the current iteration number, and epochs represents the total number of iterations;
[0102] Preferably, the number of epochs is 300;
[0103] As can be seen from the above formula, the final learning rate lrf = 0.0001, that is, when the learning rate of the anti-spy camera detection model reaches 0.0001, the training is complete.
[0104] The reasons for choosing to load pre-trained weights for model training are shown in Table 2:
[0105] Table 2 Comparison of the effects of YOLOv5s training methods
[0106]
[0107] As shown in Table 2, the applicant compared the test performance of the models trained using three different methods on the validation set of the S1 dataset. The pre-trained weights in these three methods refer to those obtained from training the YOLOv5s network on the COCO dataset. The first method, training without pre-trained weights, means randomly initializing the network weights during training. The second method, training with pre-trained weights, means initializing the network weights using the weights trained on the COCO dataset. The third method, training with pre-trained weights and freezing the network backbone, means initializing the network weights with the weights trained on the COCO dataset and then training only the weights after the network backbone. Table 2 shows that training with pre-trained weights yields the best test performance.
[0108] After the training in step S202 is completed, a first detection network is obtained, which can initially eliminate non-spy photography behavior. In order to determine whether there is spy photography behavior, the inventors summarized two important characteristics of spy photography behavior, especially spy photography mobile phones: (1) the spy photography mobile phone will be in a horizontal or vertical state to ensure the viewing angle is upright; (2) the spy photography mobile phone will remain stationary to ensure clear focus.
[0109] In view of the above-mentioned characteristics of surreptitious filming, preferably, step S200 further includes:
[0110] S203: Determination of voyeurism.
[0111] In another preferred embodiment, in step S200, as Figure 3 As shown, the step of determining whether a mobile phone in a video image to be detected is in a state of surreptitious filming according to a preset detection strategy includes the following steps:
[0112] S2031: Determine whether the mobile phone in the video image to be detected is horizontal or deviates from the horizontal by a certain angle, and whether it is vertical or deviates from the vertical by a certain angle.
[0113] In this step, after the model detects the phone, it obtains the phone's location and dimensions (x, y, w, h). x and y represent the top-left position of the phone's detection bounding box, and w and h represent the dimensions of the detection box. Since the detection box is close to the edge of the phone when it is placed horizontally or vertically, the ratio of the shorter side to the longer side of the detection box is minimized at this time. The ratio Rwh is as follows:
[0114] Rwh = min(w,h) / max(w,h)
[0115] When Rwh is less than the threshold T1, it can be preliminarily determined that the phone is in a state of covert filming; otherwise, the judgment ends. Regarding the value of T1, considering that most phones on the market have an aspect ratio of 16:9 or 16:10, meaning the ratio of the shorter side to the longer side of a typical phone is less than 0.625, the theoretical range for T1 is between 0.625 and 1.0. A T1 of 0.625 means that a small number of phones and some phones tilted at a small angle for covert filming will be missed, resulting in a high false positive rate and a low false negative rate. A T1 of 1.0 means that most tilted phones that are not being used for covert filming will be falsely detected, resulting in a low false positive rate and a high false negative rate. In balancing the false positive and false negative rates, this embodiment selects a threshold T1 range of (0.65-0.85). If T1 is less than 0.65, many phones will be missed, resulting in a high false negative rate. If T1 is greater than 0.85, many normally used phones will also be mistaken for covert filming phones, resulting in a high false negative rate.
[0116] S2032: Determine whether the mobile phone in the video image to be detected remains stationary.
[0117] In this step, the position of the mobile phone detection frame in the two consecutive frames can be compared to determine whether the mobile phone is in a state of being secretly photographed. However, due to the existence of errors, it cannot be guaranteed that the phone is truly stationary. Therefore, this embodiment uses the Intersection over Union (IOU) of the detection frames in the two consecutive frames to make the determination. The calculation formula is as follows:
[0118]
[0119] and
[0120]
[0121] Where (x1, y1, w1, h1) are the mobile phone detection box information of the previous frame; (x2, y2, w2, h2) are the mobile phone detection box information of the current frame, and the intersection area of the detection boxes is Iarea. IOU=1.0 represents theoretical complete stillness, IOU=0.0 represents significant mobile phone movement, when IOU is greater than the threshold T2 it means that the mobile phone is actually still and there is a hidden camera behavior, otherwise the judgment ends.
[0122] This embodiment takes into account factors such as hand tremors and model detection errors. It is difficult for the detection frames of two consecutive frames to overlap 100% or be completely still. Therefore, this embodiment sets an error range (0.03-0.08), that is, the IOU threshold T2 range (0.92-0.97). If the T2 value is less than 0.92, a normally used, slowly moving phone may be falsely detected as a hidden camera, resulting in a high false detection rate. If the T2 value is greater than 0.97, a hidden camera may be missed due to external factors such as model, lighting, and slight hand tremors, resulting in a high missed detection rate.
[0123] Furthermore, a phone-based surreptitious filming activity is only determined when both features are simultaneously satisfied. This invention determines the existence of such activity solely through preliminary S1 data processing and phone detection information. Compared to other methods that utilize human body information, hand information, and phone information, this model is smaller, faster, and better meets real-time requirements.
[0124] In another embodiment, before performing step S200, the following steps also need to be performed:
[0125] S101: Determine whether the video image to be detected contains a mobile phone;
[0126] In this step, if the captured video image does not contain a mobile phone, it is directly discarded, and only video images containing a mobile phone are retained as the video images to be detected.
[0127] S102: Segment the video image to be detected that contains a mobile phone to obtain a video image containing only a mobile phone;
[0128] In this step, since the acquired video images to be detected contain other interference besides the mobile phone, such as the surrounding environment and human bodies, in order to improve the detection efficiency of the model, it is necessary to segment the video images to be detected and retain only the video images containing the mobile phone.
[0129] In another embodiment, this application also provides an anti-spy phone camera device, comprising:
[0130] The acquisition unit is used to acquire the video image to be detected;
[0131] The judgment unit is used to input the acquired video image to be detected into the trained anti-spy camera detection model, and to determine whether the mobile phone in the video image to be detected is in a spy camera state according to the preset detection strategy.
[0132] In another embodiment, the determining unit includes:
[0133] The model training subunit is used to train the anti-spy camera model;
[0134] The model validation subunit is used to validate the trained anti-spy camera model.
[0135] The input sub-unit is used to input the acquired video image to be detected into the trained anti-spy camera detection model;
[0136] The first judgment subunit is used to determine whether the mobile phone in the video image to be detected is horizontal or deviates from the horizontal by a certain angle and whether it is vertical or deviates from the vertical by a certain angle.
[0137] The second judgment subunit is used to determine whether the mobile phone in the video image to be detected remains stationary.
[0138] In another embodiment, this application also provides a computer device, including:
[0139] Memory, used to store computer instructions;
[0140] A processor is used to execute computer instructions to implement the target detection method described above.
[0141] In another embodiment, such as Figure 4 As shown, this application also provides an application of a method for preventing mobile phone spying, wherein the method is applied to a commercial display screen.
[0142] The anti-spy camera detection model given in the foregoing embodiments can detect mobile phone spy camera behavior on a PC. The purpose of this embodiment is to further apply this model to commercial display screens to fill the current gap in anti-spy camera technology for commercial display screens. The specific application steps include the following:
[0143] 1. Download the compilation tool pnnx (PyTorch Neural Network Exchange) and use pnnx to convert the PyTorch version of the anti-spy camera detection model into an ncnn model (ncnn is a feedforward network framework designed for mobile devices, which runs faster on the Android platform).
[0144] 2. Use C++ to call the ncnn model, encapsulate the JNI interface, and compile it into an Android library;
[0145] 3. Package the Android APK and port it to the commercial display screen for installation and operation.
[0146] In another embodiment, this application also provides an application of an anti-spy phone camera device, wherein the device is applied to a commercial display screen.
Claims
1. A target detection method, comprising the following steps: S100: Acquire the video image to be detected; S200: Input the acquired video image to be detected into the target detection model to determine whether the video image to be detected is being secretly filmed; Step S200 includes: S203: Determining if someone is secretly filming; The surreptitious filming determination includes: S2031: determining whether the mobile phone in the video image to be detected is horizontal or deviating from the horizontal by a certain angle and whether it is vertical or deviating from the vertical by a certain angle. The determination method is as follows: after the target detection model detects the mobile phone, the location and width and height information (x, y, w, h) of the mobile phone can be obtained. x and y represent the position of the upper left point of the mobile phone detection box, w and h represent the width and height information of the mobile phone detection box, and the ratio of the short side to the long side is Rwh. When Rwh is less than the threshold T1, it can be preliminarily determined that the mobile phone is in a surreptitious filming state; otherwise, the determination ends. S2032: Determine whether the mobile phone in the video image to be detected remains stationary. The determination method is to calculate the Intersection over Union (IOU) of the mobile phone detection boxes in the two consecutive frames. If the IOU is greater than the threshold T2, it means that the mobile phone is actually stationary and there is a hidden camera behavior. Otherwise, the determination ends.
2. The method according to claim 1, wherein, In step S200, the target detection model adopts a single-stage target detection model, which includes: A network input layer, wherein the network input layer is used to input the video image to be detected; A convolutional block, which consists of a convolutional layer, a batch normalization layer, and a sigmoid activation layer; C3 block, wherein the C3 block is a residual cascaded unit composed of multiple convolutional blocks; SPFF blocks, wherein the SPFF blocks are fast pooling units consisting of 2 convolutional layers and 3 max pooling layers; An upsampling layer, wherein the upsampling layer uses the nearest neighbor method for upsampling; Feature map channel connection layer; The detection block consists of three convolutional activation units. Each convolutional activation unit processes the output detection information at one scale level. The convolutional activation unit consists of a convolutional layer and a sigmoid activation layer. After processing the detection information at three scale levels, multiple candidate boxes are obtained. Finally, the final mobile phone detection box is output through the maximum suppression method.
3. The method according to claim 1, wherein, Step S200 further includes a model training step, specifically including: S201: Collect mobile phone video images including those taken in a surreptitious state and those not taken in a surreptitious state, and divide them into training set and validation set; S202: Input the training set into the object detection model to train the model. Training is complete when the learning rate of the object detection model reaches the preset value.
4. The method according to claim 3, wherein, The model training steps employ a pre-trained weight loading method to train the model, specifically including: The dataset prepared in the data preparation steps is used as the network input. During model training, stochastic gradient descent is used to update the network weights, and the custom learning rate decay formula is as follows. lr = lr0 ((1 – x / epochs) 0.99+0.01), Where lr represents the learning rate for each iteration, lr0 represents the initial learning rate, x represents the current iteration number, and epochs represents the total number of iterations.
5. The method according to claim 1, characterized in that: Step S2031 further includes: the ratio of the short side to the long side, Rwh, is: Rwh=min(w,h) / max(w,h).
6. The method according to claim 1, characterized in that: Step S2032 further includes: The formula for calculating IOU is as follows: and Where (x1,y1,w1,h1) is the mobile phone detection box information of the previous frame; (x2,y2,w2,h2) is the mobile phone detection box information of the current frame, and the intersection area of the detection boxes is Iarea.
7. A target detection device, comprising: The acquisition unit is used to acquire the video image to be detected; The judgment unit is used to input the acquired video image to be detected into the trained target detection model to determine whether the mobile phone in the video image to be detected is in a state of surreptitious filming. The surreptitious filming determination includes: determining whether the mobile phone in the video image to be detected is horizontal or deviating from the horizontal within a certain angle, and whether it is vertical or deviating from the vertical within a certain angle. The determination method is as follows: after the target detection model detects the mobile phone, the location and width and height information (x, y, w, h) of the mobile phone can be obtained. x and y represent the position of the upper left point of the mobile phone detection box, w and h represent the width and height information of the mobile phone detection box, and the ratio of the shorter side to the longer side is Rwh. When Rwh is less than the threshold T1, it can be preliminarily determined that the mobile phone is in a surreptitious filming state; otherwise, the determination ends. The determination method is to determine whether the mobile phone in the video image to be detected remains stationary. The determination method is to calculate the intersection-union ratio (IOU) of the mobile phone detection boxes of the two frames before and after. When the IOU is greater than the threshold T2, it means that the mobile phone is actually stationary and there is surreptitious filming behavior; otherwise, the determination ends.