A lifting podium operating system based on visual algorithm to control motor lifting
By using visual algorithms and CBAM attention mechanism on the lifting podium for object detection and tracking, combined with the depth camera for distance estimation, the problem of insufficient ability to adapt to height and environment in the existing technology is solved, and intelligent lifting control is realized , improves the robustness of the model and the safety and stability of the system.
Patent Information
- Application Number
- CN202410378535.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-29
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-03-29
AI Technical Summary
The existing lifting podium cannot automatically adapt to the needs of speakers of different heights. The control method is simple, lacks intelligence, and cannot be effectively adapted in different environments. The model is not robust and generalized, and the security and stability are not high. .
Using an operating system based on vision algorithms, object detection and tracking is performed through Yol oV5 algorithm and CBAM attention mechanism, distance estimation is performed in combination with depth cameras, automatic lifting and falling, API interface is provided and deployed on intelligent host edge devices, using model cuts Branch and quantitative technology accelerates, enhancing the model's attention to important features and the security of the system.
It realizes automatic adaptation to the needs of speakers of different heights, improves the generalization ability and robustness of the model, enhances feature expression ability, ensures the safety and stability of the system, and meets the comfortable station of the speaker during the speech process. bit demand.
Smart Images

Figure CN118226961B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of a lifting podium with a motor controlled by a visual algorithm, and in particular to an operating system for a lifting podium with a motor controlled by a visual algorithm. Background Art
[0002] In modern society, the development of science and technology is constantly changing the way we live, work and learn. Many traditional tools and equipment have been replaced by more efficient, convenient and ergonomic products. Just like the podium, a facility that plays a core role in many fields such as education, business, and enterprise, its traditional product model is often designed with a fixed height, or the height of the podium is changed by physically increasing or lowering it, but the speaker needs to cooperate in advance before the speech, which cannot meet the needs of speakers of different heights. There is no intelligent equipment that can automatically rise and fall to the corresponding height of the speaker. At the same time, its hardware facilities are relatively simple, and usually require additional external equipment to meet more complex needs, such as microphones, demonstration equipment, etc. This not only increases the complexity of the speaker's use, but also the external wiring may cause inconvenience in use.
[0003] At present, the lifting podium is only controlled by ordinary button connections. Obviously, this control method has the following problems: 1. There is no selection of people with different backgrounds, which cannot improve the model's adaptability to different environments; there is no selection of people with different genders, ages, body shapes and clothing, which makes it impossible to enhance the generalization ability of the model; there is no data enhancement, which makes it impossible to expand the data set and improve the robustness of the model; at the same time, the data set is not divided into training set, validation set and test set, which cannot ensure that the samples of each category are balanced in the data set, and cannot avoid the model from being biased towards a certain category; CBAM is not used as the attention mechanism, and the CBAM module is not embedded between the feature extraction layer and the neck network of YoloV5, so that the CBAM module cannot use the attention mechanism in the two dimensions of channel and space, which makes it impossible to enhance the model's attention to important features, and to a certain extent reduces the expression ability of features, and at the same time cannot take into account the adjustment of small target detection strategy and changes in anchor frame design.
[0004] 2. This application does not implement GPU acceleration of the model, and does not use model pruning and quantization techniques to reduce model size and computational complexity; it does not provide an API interface, which makes it extremely inconvenient for other systems to call tracking functions; it does not deploy the tracking solution on the edge device of the smart host, and cannot guarantee the security and stability of the system; it does not use a depth camera for distance screening and use the vertical coordinate information of the target bounding box to estimate the rising range of the podium. It does not rotate the depth camera and correspond the depth image with the plane image, and cannot implement an effective distance estimation solution. Summary of the invention
[0005] In view of the above-mentioned technical deficiencies, the purpose of this application is to provide a lifting podium operating system that controls the lifting of the motor based on a visual algorithm.
[0006] To solve the above technical problems, the present application adopts the following technical solutions: The present application provides a lifting podium operating system based on a visual algorithm to control the lifting of a motor, including: a target detection algorithm configuration module, which is used to use the YoloV5 algorithm as the target detection algorithm, introduce an attention mechanism and a special small target detection strategy, extract pre-collected and integrated image data, complete the training of the target detection algorithm, and use the target detection algorithm to complete the configuration of the target tracking solution.
[0007] The manual fine-tuning detection module is used to detect whether the manual fine-tuning lock is triggered. When the manual fine-tuning lock is triggered, the motor adjusts the lifting podium according to the signal transmitted by the pin serial port; when the manual fine-tuning lock is not triggered, the data acquisition module is entered.
[0008] The data acquisition module is used to use the target detection algorithm through the distance estimation scheme of the depth camera, and obtain the coordinates of the bounding box of each detected target after distance screening, thereby obtaining the face coordinates and face depth information.
[0009] The data analysis module is used to determine whether a person has changed in front of the lifting podium through a target tracking solution. When a person has changed in front of the lifting podium, the data acquisition module is entered to obtain the face coordinates and face depth information again. When the person has not changed in front of the lifting podium, a logical judgment is made as to whether the lifting podium should rise or fall, and then the rising or falling range of the lifting podium is obtained, so as to control the lifting podium to rise or fall through the motor.
[0010] Preferably, the pre-collected and integrated image data is specifically collected as follows: using a camera equipped with a depth camera as a data acquisition device, selecting environments with various backgrounds in various time periods, and collecting images of people of various genders, ages, body shapes and clothing in various scenes; then integrating the collected character images, first performing data cleaning, and then using annotation tools to annotate the targets in the images, including bounding boxes, categories and difficulty levels, to complete data enhancement; finally, dividing the data set into a training set, a validation set and a test set in a ratio of 7:2:1 or 8:1:1.
[0011] Preferably, the training of the target detection algorithm is completed, and the specific training process is as follows: A1. CBAM is used as the attention mechanism, and the CBAM module is embedded between the feature extraction layer and the neck network of YoloV5.
[0012] A2. Use data augmentation technology such as image scaling and cropping to increase the training samples of small targets, and then fuse feature maps at different levels to capture the features of both large and small targets.
[0013] A3. Adjust the size and proportion of the anchor box according to the shape and size of the small target.
[0014] A4. Modify the loss function and give higher weights to small targets to increase the model's attention to small targets.
[0015] A5. Use the Adam optimizer and set the initial learning rate to 0.001, the batch size to 32, and the number of iterations to 100 epochs.
[0016] A6. Use accuracy, average precision AP, and recall to evaluate model performance and complete the training of the target detection algorithm.
[0017] Preferably, the target tracking scheme is configured using a target detection algorithm, and the specific configuration process is as follows: B1. Target detection is performed using a target detection algorithm at the target position in the initial frame preset by the user, thereby completing the target initialization setting.
[0018] B2. Use the FaceNet model to extract the features of the target and generate a highly discriminative feature vector; and use an online learning strategy to update the target model in real time to adapt to changes in the target's appearance.
[0019] B3. Based on the correlation filtering tracking algorithm MOSSE, the deep learning features are combined to improve the tracking performance; and the scale-adaptive search area is used to achieve an efficient motion model.
[0020] B4. When detecting occlusion events, use optical flow or depth information and use the prediction model to maintain tracking during occlusion.
[0021] B5. Use the regression model to refine the bounding box position and use the IoU loss function to optimize the accuracy of the bounding box.
[0022] B6. Use OTB, UAV123, and VOT as standard data sets for testing, and perform customized training and testing based on pre-collected data for specific scenarios. Use accuracy, success rate, and robustness as indicators to evaluate tracking performance and compare the performance of different algorithms in similar scenarios. At the same time, test the real-time performance of the system on various hardware platforms to optimize the algorithm and model.
[0023] B7. Integrate the tracking system into the existing intelligent recognition system, provide an API interface, and then deploy it on the intelligent host edge device to complete the configuration of the target tracking solution.
[0024] Preferably, the motor adjusts the lifting podium according to the signal transmitted by the pin serial port, and the specific adjustment process is as follows: the rising and falling commands are written into the key control system, and the pins of the host are connected to the keys to realize the communication of the hardware interface; when the user manually controls the lifting and lowering of the podium through the keys, the motor is adjusted according to the signal transmitted by the pin serial port.
[0025] Preferably, the facial coordinates and facial depth information are obtained as follows: C1. When using a depth camera to measure the distance between a target object and the camera, the depth camera uses structured light technology to measure the depth, projects a light spot onto the surface of the object through an infrared dot matrix projector, and then uses an infrared camera to capture the image of each light spot, and then analyzes the displacement in the light spot image to obtain the position coordinates and depth information of the object.
[0026] C2. According to the preset distance threshold, exclude objects whose distance from the camera exceeds this threshold to reduce the interference of multiple targets and complete the distance screening; and use the target detection algorithm to identify and locate the podium, obtain the coordinates of the bounding box of each detected target, among which the vertical coordinate of the upper edge is obtained first; thereby obtaining the face coordinates and face depth information.
[0027] Preferably, the rising or falling range of the lifting podium is obtained by the following specific acquisition process: the vertical coordinate of the bounding box is mapped to the actual physical height according to the camera angle and the relative position of the target and the camera; the height change is monitored by the change of the vertical coordinate between consecutive frames, thereby obtaining the rising or falling range of the lifting podium.
[0028] Preferably, the lifting and lowering of the lifting platform is controlled by a motor, and the specific control process is as follows: after the motor receives the signal of the lifting or lowering range of the lifting platform, the lifting platform is driven to rise or fall through the motor driver and the controller.
[0029] The beneficial effects of the present application are: 1. The present application provides a lifting podium operating system that controls the lifting of a motor based on a visual algorithm. First, a target detection model is trained and a CBAM module is embedded. At the same time, the small target detection strategy is adjusted, and then the anchor frame design is changed, thereby enhancing the model's attention to important features, thereby improving the expressiveness of features; by designing a target tracking scheme, it is determined whether a person should be replaced in front of the lifting podium, and finally the lifting range of the lifting podium is obtained. The lifting podium is driven by the motor driver and controller to lift and lower. The whole process does not require any additional equipment, which meets the needs of the speaker during the speech. The overall shape is ergonomic and can meet the speaker's comfortable standing position during the speech.
[0030] 2. This application selects image data with different backgrounds to improve the model's ability to adapt to different environments; it also selects image data of people of different genders, ages, body shapes, and clothing to enhance the generalization ability of the model; it performs data enhancement on the image data to expand the data set and improve the robustness of the model; and it divides the data set into training set, validation set, and test set to ensure that samples of each category are balanced in the data set to avoid the model being biased towards a certain category.
[0031] 3. This application adopts CBAM as the attention mechanism and embeds the CBAM module between the feature extraction layer and the neck network of YoloV5. The CBAM module enhances the model's attention to important features and improves the expressiveness of features through the attention mechanism in two dimensions: channel and space, while taking into account the adjustment of small target detection strategy and changes in anchor frame design.
[0032] 4. This application implements GPU acceleration of the model, uses model pruning and quantization techniques to reduce model size and computational complexity; and provides an API interface to facilitate other systems to call tracking functions; at the same time, the tracking and detection solution is deployed on the edge device of the smart host to ensure the security and stability of the system.
[0033] 5. This application uses a depth camera for distance screening and utilizes the vertical coordinate information of the target bounding box to estimate the rising range of the podium. By rotating the depth camera and matching the depth image with the plane image, an effective distance estimation scheme is implemented. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0035] Figure 1 This is a schematic diagram of the system structure connection for this application.
[0036] Figure 2 This is a flowchart of the working steps for applying for a lifting podium.
[0037] Figure 3 This is the floor plan of the USB foot switch.
[0038] Figure 4 The electrical schematic diagram.
[0039] Figure 5 Schematic diagram of the circuit board connecting the locking and fine-tuning buttons to the smart host.
[0040] Figure 6 PCB diagram for connecting the locking and fine-tuning buttons to the smart host. DETAILED DESCRIPTION
[0041] 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.
[0042] See also Figure 1 As shown, the present application provides, in a first aspect, a lifting podium operating system for controlling the lifting of a motor based on a visual algorithm, including a target detection algorithm configuration module, a manual fine-tuning detection module, a data acquisition module, and a data analysis module.
[0043] The target detection algorithm configuration module is used to use the YoloV5 algorithm as the target detection algorithm, introduce the attention mechanism and a special small target detection strategy, extract the pre-collected and integrated image data, complete the training of the target detection algorithm, and use the target detection algorithm to complete the configuration of the target tracking solution.
[0044] In a specific example, the pre-collected and integrated image data is collected in the following manner: a camera equipped with a depth camera is used as a data acquisition device, an environment with various backgrounds is selected in each time period, and images of people of various genders, ages, body shapes, and clothing in each scene are collected; the collected character images are then integrated, firstly data cleaning is performed, and then annotating tools are used to annotate the targets in the image, including bounding boxes, categories, and difficulty levels, to complete data enhancement; finally, the data set is divided into a training set, a validation set, and a test set in a ratio of 7:2:1 or 8:1:1.
[0045] It should be noted that various time periods include day, night, dusk, etc.
[0046] It should be noted that the background environments include indoor, outdoor, urban, rural, etc.
[0047] It should be noted that it includes bounding boxes, categories, and difficulty levels such as small objects, occlusions, etc.
[0048] This application selects image data with different backgrounds to improve the model's adaptability to different environments; it also selects image data of people of different genders, ages, body shapes, and clothing to enhance the generalization ability of the model; it performs data enhancement on the image data to expand the data set and improve the robustness of the model; and it divides the data set into training, validation, and test sets to ensure that samples of each category are balanced in the data set to avoid the model being biased towards a certain category.
[0049] In a specific example, the training of the target detection algorithm is completed, and the specific training process is as follows: A1. CBAM is used as the attention mechanism, and the CBAM module is embedded between the feature extraction layer and the neck network of YoloV5.
[0050] A2. Use data augmentation technology such as image scaling and cropping to increase the training samples of small targets, and then fuse feature maps at different levels to capture the features of both large and small targets.
[0051] A3. Adjust the size and proportion of the anchor box according to the shape and size of the small target.
[0052] A4. Modify the loss function and give higher weights to small targets to increase the model's attention to small targets.
[0053] A5. Use the Adam optimizer and set the initial learning rate to 0.001, the batch size to 32, and the number of iterations to 100 epochs.
[0054] A6. Use accuracy, average precision AP, and recall to evaluate model performance and complete the training of the target detection algorithm.
[0055] It should be noted that CBAM stands for Convolutional Black Attention Module, which is an attention module that integrates channel and spatial attention.
[0056] It should be noted that the Adam optimizer can adjust different learning rates for each different parameter, update frequently changing parameters with smaller step sizes, and update sparse parameters with larger step sizes.
[0057] This application adopts CBAM as the attention mechanism and embeds the CBAM module between the feature extraction layer and the neck network of YoloV5. The CBAM module enhances the model's attention to important features and improves the expressiveness of features through the attention mechanism in two dimensions: channel and space, while taking into account the adjustment of small target detection strategy and the change of anchor frame design.
[0058] In a specific example, the target tracking solution is configured using a target detection algorithm, and the specific configuration process is as follows: B1. Target detection is performed using a target detection algorithm at the target position in the initial frame preset by the user, thereby completing the target initialization setting.
[0059] B2. Use the FaceNet model to extract the features of the target and generate a highly discriminative feature vector; and use an online learning strategy to update the target model in real time to adapt to changes in the target's appearance.
[0060] B3. Based on the correlation filtering tracking algorithm MOSSE, the deep learning features are combined to improve the tracking performance; and the scale-adaptive search area is used to achieve an efficient motion model.
[0061] B4. When detecting occlusion events, use optical flow or depth information and use the prediction model to maintain tracking during occlusion.
[0062] B5. Use the regression model to refine the bounding box position and use the IoU loss function to optimize the accuracy of the bounding box.
[0063] B6. Use OTB, UAV123, and VOT as standard data sets for testing, and perform customized training and testing based on pre-collected data for specific scenarios. Use accuracy, success rate, and robustness as indicators to evaluate tracking performance and compare the performance of different algorithms in similar scenarios. At the same time, test the real-time performance of the system on various hardware platforms to optimize the algorithm and model.
[0064] B7. Integrate the tracking system into the existing intelligent recognition system, provide an API interface, and then deploy it on the intelligent host edge device to complete the configuration of the target tracking solution.
[0065] It should be noted that the FaceNet model is a classic face recognition model, and its source code can be obtained on the Internet. It mainly maps face images into a multidimensional space and represents the similarity of faces through spatial distance.
[0066] It should be noted that the scale-adaptive search region is used to implement an efficient motion model, such as a search based on particle filtering or mean shift.
[0067] It should be noted that when using quantitative forecasting methods for forecasting, the most important task is to establish a forecasting mathematical model, which describes the quantitative relationship between things in mathematical language or formulas, and uses it as the direct basis for calculating the forecast value when forecasting.
[0068] It should be noted that the regression model is a mathematical model that quantitatively describes statistical relationships.
[0069] It should be noted that the IoU loss function is one of the commonly used indicators in target detection algorithms.
[0070] It should be noted that the motion model, prediction model, regression model and IoU loss function are all models that can be downloaded from the Internet.
[0071] This application implements GPU acceleration of the model, uses model pruning and quantization techniques to reduce model size and computational complexity; and provides an API interface to facilitate other systems to call tracking functions; at the same time, the tracking and detection solution is deployed on the edge device of the smart host to ensure the security and stability of the system.
[0072] The manual fine-tuning detection module is used to detect whether the manual fine-tuning lock is triggered. When the manual fine-tuning lock is triggered, the motor adjusts the lifting podium according to the signal transmitted by the pin serial port; when the manual fine-tuning lock is not triggered, the data acquisition module is entered.
[0073] In a specific example, the motor adjusts the lifting podium according to the signal transmitted by the pin serial port, and the specific adjustment process is as follows: the rising and falling commands are written into the key control system, and the pins of the host are connected to the keys to realize the communication of the hardware interface; when the user manually controls the lifting and lowering of the podium through the keys, the motor is adjusted according to the signal transmitted by the pin serial port.
[0074] The data acquisition module is used to use the target detection algorithm through the distance estimation scheme of the depth camera, and obtain the coordinates of the bounding box of each detected target after distance screening, thereby obtaining the face coordinates and face depth information.
[0075] In a specific example, the facial coordinates and facial depth information are obtained as follows: C1. When using a depth camera to measure the distance between a target object and the camera, the depth camera uses structured light technology to measure the depth, projects a light spot onto the surface of the object through an infrared dot matrix projector, and then uses an infrared camera to capture the image of each light spot, and then analyzes the displacement in the light spot image to obtain the position coordinates and depth information of the object.
[0076] C2. According to the preset distance threshold, exclude objects whose distance from the camera exceeds this threshold to reduce the interference of multiple targets and complete the distance screening; and use the target detection algorithm to identify and locate the podium, obtain the coordinates of the bounding box of each detected target, among which the vertical coordinate of the upper edge is obtained first; thereby obtaining the face coordinates and face depth information.
[0077] It should be noted that the preset distance threshold is, for example, 1 meter, 2 meters, etc.
[0078] The data analysis module is used to determine whether a person has changed in front of the lifting podium through a target tracking solution. When a person has changed in front of the lifting podium, the data acquisition module is entered to obtain the face coordinates and face depth information again. When the person has not changed in front of the lifting podium, a logical judgment is made as to whether the lifting podium should rise or fall, and then the rising or falling range of the lifting podium is obtained, so as to control the lifting podium to rise or fall through the motor.
[0079] In a specific example, the rising or falling range of the lifting podium is obtained by the following specific acquisition process: the vertical coordinate of the bounding box is mapped to the actual physical height according to the camera angle and the relative position of the target and the camera; the height change is monitored by the change of the vertical coordinate between consecutive frames, thereby obtaining the rising or falling range of the lifting podium.
[0080] This application uses a depth camera for distance screening and utilizes the vertical coordinate information of the target bounding box to estimate the rising range of the podium. By rotating the depth camera and matching the depth image with the plane image, an effective distance estimation solution is implemented.
[0081] In a specific example, the lifting platform is lifted and lowered by controlling the lifting platform with a motor, and the specific control process is as follows: after the motor receives the signal of the lifting platform's rising or falling range, the lifting platform is driven to rise or fall through the motor driver and controller.
[0082] It should be noted that after the automatic lifting is completed, the user can still lock the automatic lifting button, switch to manual lifting mode, and perform manual fine-tuning.
[0083] Depend on Figure 3 As shown, the lifting podium in this application also provides a control solution for the PPT page turning keys, using a USB foot switch industrial control circuit board to simulate the mouse buttons to realize the PPT page turning function; the internal design of the circuit board has the ability to simulate mouse movement and scroll wheel functions, and supports 8 switches at the same time, allowing users to customize key values and combination keys; the commands corresponding to the up and down keys of the computer are soldered to the two switches of the circuit board, and the up and down pages of the PPT are realized through the foot switch.
[0084] Depend on Figure 4 As shown, the present application also provides an electrical intelligent integration solution. After the intelligent host is started, it retrieves the camera input information, performs algorithm processing, and converts the signal to drive the motor; drives the central axis of the motor to drive the outer sleeve and the table top to rise and fall; adds multiple AC 220V to DC power supplies inside the column to power various devices; adds a delay relay to prevent the podium from falling suddenly during the power-on process; DC power of each voltage level is used to power the motor, motor controller, touch screen computer, bar display, router, and delay relay equipment.
[0085] The present application provides an operating system for a lifting podium that controls the lifting of a motor based on a visual algorithm. First, a target detection model is trained and a CBAM module is embedded. At the same time, the small target detection strategy is adjusted, and then the anchor frame design is changed to enhance the model's attention to important features, thereby improving the expressiveness of features. By designing a target tracking scheme, it is determined whether a person should be replaced in front of the lifting podium, and finally the lifting range of the lifting podium is obtained. The lifting podium is driven to rise and fall by the motor driver and controller. The whole process does not require any additional equipment, which meets the needs of the speaker during the speech. The overall shape is ergonomic and can meet the speaker's comfortable standing position during the speech.
[0086] The above contents are merely examples and explanations of the concept of the present application. The technical personnel in the relevant technical field may make various modifications or additions to the described specific embodiments or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined in this specification, they should all fall within the scope of protection of the present application.
Claims
1. A lifting podium operating system based on visual algorithm to control the motor lifting. It is characterized in that include: The target detection algorithm configuration module is used to use the YoloV5 algorithm as the target detection algorithm, introduce the attention mechanism and a special small target detection strategy, extract the pre-collected and integrated image data, complete the training of the target detection algorithm, and use the target detection algorithm to complete the configuration of the target tracking solution; The specific collection process of the pre-collected and integrated image data is as follows: Using a camera equipped with a depth camera as a data collection device, select environments with various backgrounds at different time periods to collect images of people of different genders, ages, body shapes, and clothing in each scene; then integrate the collected person images, first perform data cleaning, and then use annotation tools to annotate the objects in the image, including bounding boxes, categories, and difficulty levels, to complete data enhancement; finally, divide the data set into training set, validation set, and test set, with a ratio of 7:2:1 or 8:1:1; The training of the target detection algorithm is completed, and the specific training process is as follows: A1. Use CBAM as the attention mechanism and embed the CBAM module between the feature extraction layer and the neck network of YoloV5; A2. Use image scaling and cropping data enhancement technology to increase the training samples of small targets, and then fuse feature maps at different levels to capture the features of large and small targets at the same time; A3. Adjust the size and proportion of the anchor box according to the shape and size of the small target; A4. Modify the loss function and give higher weights to small targets to increase the model's attention to small targets. A5. Use the Adam optimizer and set the initial learning rate to 0.001, the batch size to 32, and the number of iterations to 100 epochs. A6. Use accuracy, average precision AP, and recall to evaluate model performance, thereby completing the training of the target detection algorithm. The target detection algorithm is used to complete the configuration of the target tracking solution. The specific configuration process is as follows: B1. Use the target detection algorithm to detect the target position in the initial frame preset by the user, thereby completing the initialization setting of the target; B2. Use the FaceNet model to extract the features of the target and generate a highly discriminative feature vector; and use an online learning strategy to update the target model in real time to adapt to changes in the target's appearance; B3. Based on the correlation filtering tracking algorithm MOSSE, combined with deep learning features to improve tracking performance; and use scale-adaptive search areas to achieve efficient motion models; B4. When detecting occlusion events, use optical flow or depth information and use prediction models to maintain tracking during occlusion. B5. Use the regression model to refine the bounding box position and use the IoU loss function to optimize the accuracy of the bounding box; B6. Use OTB, UAV123, and VOT as standard data sets for testing, and perform customized training and testing based on pre-collected data for specific scenarios; use accuracy, success rate, and robustness as indicators to evaluate tracking performance, and compare the performance of different algorithms in similar scenarios; test the real-time performance of the system on various hardware platforms, and optimize the algorithm and model; B7. Integrate the tracking system into the existing intelligent recognition system, provide an API interface, and then deploy it on the intelligent host edge device to complete the configuration of the target tracking solution; The manual fine-tuning detection module is used to detect whether the manual fine-tuning lock is triggered. When the manual fine-tuning lock is triggered, the motor adjusts the lifting platform according to the signal transmitted by the pin serial port; when the manual fine-tuning lock is not triggered, the data acquisition module is entered; The motor adjusts the lifting platform according to the signal transmitted by the pin serial port. The specific adjustment process is as follows: Write the commands of rising and falling into the key control system, and connect the pins of the host to the keys to realize the communication of the hardware interface; when the user manually controls the rise and fall of the podium through the keys, the motor is adjusted according to the signal transmitted by the pin serial port; A data acquisition module is used to use the target detection algorithm through the distance estimation scheme of the depth camera, and after performing distance screening, obtain the coordinates of the bounding box of each detected target to obtain the face coordinates and face depth information; The specific contents of obtaining face coordinates and face depth information are as follows: C1. When using a depth camera to measure the distance between a target object and the camera, the depth camera uses structured light technology to measure the depth. It projects light spots onto the surface of the object through an infrared dot matrix projector, and then uses an infrared camera to capture the image of each light spot. It then analyzes the displacement in the light spot image to obtain the position coordinates and depth information of the object. C2. According to the preset distance threshold, exclude objects whose distance from the camera exceeds this threshold, reduce the interference of multiple targets, and then complete the distance screening; and use the target detection algorithm to identify and locate the podium, obtain the coordinates of the bounding box of each detected target, wherein the vertical coordinate of the upper edge is obtained first; thereby obtaining the face coordinates and face depth information; The data analysis module is used to determine whether a person has been replaced in front of the current lifting platform through a target tracking solution. When a person has been replaced in front of the lifting platform, the data acquisition module is entered to obtain the face coordinates and face depth information again. When a person has not been replaced in front of the lifting platform, a logical judgment is made as to whether the lifting platform should rise or fall, and then the rising or falling range of the lifting platform is obtained, so as to control the lifting platform to rise or fall through the motor; The specific process of obtaining the rising or falling range of the lifting podium is as follows: According to the camera angle and the relative position of the target and the camera, the ordinate of the bounding box is mapped to the actual physical height; the change of the ordinate between consecutive frames is used to monitor the height change, thereby obtaining the rise or fall range of the lifting platform; The lifting and lowering of the lecture platform is controlled by a motor, and the specific control process is as follows: After the motor receives the signal of the rising or falling range of the lifting podium, it drives the lifting podium to rise or fall through the motor driver and the controller.
