Intelligent airflow control system of air curtain machine based on computer vision

Abstract

The invention relates to the technical field of air curtain machine control, in particular to an air curtain machine intelligent airflow control system based on computer vision, which comprises a sensing module, a data processing module, an execution module, an environment sensing module and a control module, the motion trail identification module is used for identifying a real-time motion trail of a motion target in the real-time video data through a computer vision algorithm and predicting a future motion trail of the motion target through a prediction algorithm; compared with the prior art that a passive control scheme adopting fixed wind speed or simple infrared triggering has the defects that energy consumption is high and airflow is stiff and abrupt, an active pre-control strategy based on computer vision prediction is adopted, the pedestrian trajectory is accurately predicted, and a matched airflow barrier is established in advance, so that the pedestrian trajectory is accurately predicted. Effective combination of non-inductive passing experience and on-demand energy supply is realized, and the system has the advantages of remarkably improving comfort and greatly reducing ineffective energy consumption.

Description

Technical Field

[0001] This invention relates to the field of air curtain machine control technology, and in particular to an intelligent airflow control system for air curtain machines based on computer vision. Background Technology

[0002] Air curtains, as a common environmental isolation device, are widely used at the entrances and exits of shopping malls, warehouses, office buildings, and other places. Their core function is to effectively block the free exchange of indoor and outdoor air through a downward-spraying airflow barrier, thereby maintaining indoor temperature and preventing dust and insects from entering. Traditional air curtains operate in a relatively simple and direct manner, mostly relying on continuous operation or triggering by basic sensors.

[0003] Currently, most mainstream air curtain machines use fixed-speed operation or simple passive sensing control. Fixed-speed mode means the equipment operates at a constant power regardless of whether someone is at the door or the temperature difference, resulting in significant energy waste. While passive control schemes based on infrared sensors can activate or adjust the fan speed after detecting human movement, they are essentially a delayed "reactive" control with significant latency. The fan only accelerates after a pedestrian has entered the sensing zone, causing the pedestrian's front half to feel the strong wind, resulting in poor comfort; after the pedestrian leaves, the fan cannot decelerate in time, still causing energy waste. Furthermore, existing technology cannot identify the number of pedestrians, their walking speed, intentions, or individual differences (such as whether they are carrying large items, or whether they are children or adults), let alone predict their future movement trajectory. Therefore, in complex and realistic application scenarios, its control effect appears rather mechanical and inefficient, with significant shortcomings in energy saving, comfort, and intelligence.

[0004] To address the aforementioned issues, this invention proposes an intelligent airflow control system for air curtain machines based on computer vision. Summary of the Invention

[0005] To overcome the problems mentioned in the background art, the present invention proposes an intelligent airflow control system for air curtain machines based on computer vision.

[0006] The technical solution of this invention is: an intelligent airflow control system for an air curtain machine based on computer vision, comprising:

[0007] The sensing module includes an image acquisition device located near the area covered by the air curtain machine, used to acquire real-time video data containing moving targets;

[0008] The data processing module is used to identify the real-time motion trajectory of moving targets in real-time video data through computer vision algorithms, and to predict the future motion trajectory of moving targets through prediction algorithms.

[0009] The execution module includes an air curtain machine, which includes an adjustable speed fan unit and an adjustable angle air guide plate unit;

[0010] The environmental sensing module is used to acquire information on the indoor and outdoor temperature difference in the area covered by the air curtain machine;

[0011] The control module is used to generate control commands for controlling the fan unit and the air guide plate unit based on the real-time motion trajectory and the future motion trajectory.

[0012] Preferably, when the data processing module identifies the real-time motion trajectory of a moving target in real-time video data using computer vision algorithms, it includes the following steps:

[0013] S11: Extract the current frame image from the real-time video stream acquired by the image acquisition device at a preset frame rate;

[0014] S12: Input the current frame image into the pre-trained target detection neural network model to obtain the initial detection result set of all moving targets in the frame;

[0015] S13: Associate the detection results of the current frame with the established set of tracked targets and assign an identity ID to each detected target;

[0016] S14: For a successfully associated tracking target, based on its current observations and existing motion history, a state estimation algorithm is used to update its motion state vector, which contains position and velocity information.

[0017] S15: Based on the updated motion state vector, generate trajectory information for each moving target at timestamp t.

[0018] Preferably, when the data processing module predicts the future trajectory of a moving target using a prediction algorithm, it includes the following steps:

[0019] S21: Trajectory sequence preprocessing, to obtain the smoothed state sequence of the moving target in the most recent T historical time steps, i.e., the historical trajectory sequence;

[0020] S22: Trajectory feature encoding, which inputs the historical trajectory sequence into the encoder network, extracts its implicit spatiotemporal features, and outputs a feature vector of fixed dimensions;

[0021] S23: Future trajectory decoding: The feature vector is input into the decoder network, and the decoder network recursively predicts the future target state to generate a set of predicted trajectories;

[0022] S24: Optimal trajectory selection: Select the optimal predicted trajectory from the predicted trajectory set according to preset rules.

[0023] Preferably, the data processing module also includes, during operation, coarse-grained attribute analysis of the moving target that does not involve personal identification, specifically including:

[0024] A11: Infer the thickness of clothing based on the target's contour information;

[0025] A12: Identify whether a moving target is carrying large items;

[0026] A13: Based on the height and gait characteristics of the movement target, they are classified into the age groups of children, adults, and the elderly.

[0027] Preferably, when the data processing module infers the clothing thickness based on the target's contour information, it specifically includes:

[0028] S31: Obtain the image region and trajectory information of the moving target. The image region is defined by the bounding box provided by the target detection and tracking module, and the trajectory information includes the position sequence in consecutive frames.

[0029] S32: Extract the binary contour mask of the moving target from the image region;

[0030] S33: Calculate the aspect ratio and pixel area of ​​the contour mask;

[0031] S34: The aspect ratio and pixel area are fused with the current outdoor temperature value obtained from the environmental sensor to form a feature vector;

[0032] S35: Input the fused feature vector into the classifier, and output the clothing classification label representing the thickness of the clothing through the classifier. The clothing classification label includes light and thin and thick.

[0033] Preferably, when performing attribute analysis that does not involve personal identification in parallel based on image region and trajectory information, the method also includes inferring the age group attribute of the moving target, specifically:

[0034] S41: Calculate the average speed and gait period characteristics of a moving target based on trajectory information;

[0035] S42: Based on the image region, estimate the pixel height of the moving target, and combine the camera intrinsic parameters to estimate the absolute height of the moving target;

[0036] S43: Input the estimated absolute height, average speed and gait features into the classifier, and output the age classification label representing the age group. The age classification label includes children, adults and the elderly.

[0037] Preferably, the data processing module is also used to identify preset behaviors based on the real-time motion trajectory of the moving target. The preset behaviors include staying, lingering, and suddenly turning back in the area covered by the air curtain machine.

[0038] When lingering and loitering behavior is detected, the control module generates instructions to reduce the wind speed in the corresponding area and switch to a gentle mode.

[0039] When a sudden reversal is detected, the control module generates a command to cancel the pre-loaded acceleration control of the corresponding moving target.

[0040] Preferably, the control module, when in operation, specifically includes:

[0041] S51: Continuously receives real-time motion trajectory of moving targets, predicted future motion trajectory and attribute analysis results from the data processing module, and obtains indoor and outdoor temperature difference data from the environmental perception module;

[0042] S52: Determine the current operating mode that the system should enter. Operating modes include standby mode, single-target pre-action mode, multi-target collaborative mode, and special behavior mode.

[0043] S53: Generate control commands containing the target wind speed value and the target wind deflector angle based on the determined working mode;

[0044] S54: Send control commands to the execution module; for commands based on predicted trajectories, perform timing calculations to achieve early triggering.

[0045] Preferably, when the control module determines that the current system has entered the single-target pre-action mode, the specific workflow includes:

[0046] S61: When it is predicted that a single moving target will enter the air curtain machine coverage area within a preset time threshold, calculate its expected arrival location and time based on the predicted trajectory.

[0047] S62: Based on the analysis results of the moving target's attributes and the indoor-outdoor temperature difference, fine-tune the target wind speed and the angle of the wind deflector.

[0048] S63: Based on the estimated arrival time, generate an acceleration command one system delay time in advance to ensure that the airflow is stably established before the moving target arrives.

[0049] Preferably, when the control module determines that the current system has entered the multi-target pre-action mode, the specific workflow includes:

[0050] S71: When there are multiple predicted trajectories of moving targets at the same time and the trajectories intersect within the coverage area of ​​the air curtain machine, calculate the smallest continuous area that can cover all relevant trajectories.

[0051] S72: The generated control commands are used to control the fan unit to rise to a wind speed that can cover the entire area and to control the air guide plate unit to swing.

[0052] Preferably, when the control module determines that the current system has entered a special behavior mode, it specifically includes:

[0053] A21: Dwelling Response Sub-mode. When a moving target is detected to have stayed in the area covered by the air curtain machine for more than a time threshold, an instruction is generated to reduce the wind speed in the local area where the target is located.

[0054] A22: Turnback Cancellation Sub-mode. When a moving target is detected to suddenly turn back, an instruction is generated to cancel the target's pre-loaded acceleration instruction and restore the wind speed control parameters to the state before the turnback.

[0055] Preferably, when the control module generates control commands containing the target wind speed value and the target wind deflector angle according to the determined working mode, it also includes a wind direction compensation energy-saving strategy based on the indoor and outdoor temperature difference, specifically:

[0056] When no moving target is detected, the control fan unit operates at a lower base wind speed, and the control air guide plate unit tilts towards the outdoor side at a compensation angle to achieve better airtightness isolation with low energy consumption.

[0057] When a moving target is detected, the air guide plate unit is controlled to return to the optimal isolation angle.

[0058] The beneficial effects of this invention are:

[0059] 1. Compared with the passive control schemes of existing technologies that use fixed wind speed or simple infrared triggering, which have the disadvantages of high energy consumption and abrupt and harsh airflow, this invention adopts an active pre-control strategy based on computer vision prediction. By accurately predicting pedestrian trajectories and establishing matching airflow barriers in advance, it achieves an effective combination of seamless passage experience and on-demand energy supply, which has the advantages of significantly improving comfort and greatly reducing ineffective energy consumption.

[0060] 2. Compared with existing technologies that cannot perceive individual differences among pedestrians and adopt a one-size-fits-all approach to airflow intensity, which has the disadvantages of being unfriendly to vulnerable groups such as children and the elderly or failing to isolate large items, this invention introduces anonymized attribute analysis capabilities. By intelligently identifying age groups, clothing thickness, and carried items, it can personalize the wind force, which has the advantages of reflecting humanistic care and enhancing the reliability of environmental isolation.

[0061] 3. Compared with existing technologies, which lack a mechanism to deal with complex human behaviors and have the disadvantages of causing continuous strong winds or energy waste due to people lingering at the door or temporarily turning back, this invention has refined behavior recognition and scenario mode switching functions. It can automatically deal with special scenarios such as lingering and wandering and promptly cancel invalid commands. It has the advantages of significantly improving the applicability of real scenarios and further optimizing energy efficiency.

[0062] 4. Compared with the shortcomings of existing technologies, such as control lag or over-response when multiple targets pass through, resulting in airflow chaos or energy waste, this invention adopts a multi-target cooperative control mode. By calculating the overall coverage area and generating a unified and optimized sweeping airflow, it achieves efficient management of group passage scenarios and has the advantages of improving isolation efficiency and comfort under dense crowds.

[0063] 5. Compared with existing technologies that can only adjust wind speed but cannot optimize wind direction, and have the disadvantage of difficulty in balancing energy consumption and heat preservation effect during standby, this invention proposes a wind direction compensation energy-saving strategy. During idle periods, the air tightness is enhanced by adjusting the angle of the air guide plate rather than simply maintaining the air volume, thus achieving a balance between energy saving and basic performance. It has the advantage of maintaining excellent isolation effect even at low power consumption. Attached Figure Description

[0064] Figure 1 The diagram shown is a schematic representation of the intelligent airflow control system for an air curtain machine based on computer vision according to the present invention.

[0065] Figure 2 The diagram shown is a second three-dimensional structural schematic of the intelligent airflow control system for air curtain machines based on computer vision according to the present invention. Detailed Implementation

[0066] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0067] Please see Figures 1-2 The present invention provides an embodiment of an intelligent airflow control system for an air curtain machine based on computer vision, comprising:

[0068] The sensing module includes an image acquisition device located near the area covered by the air curtain machine, used to acquire real-time video data containing moving targets;

[0069] The data processing module is used to identify the real-time motion trajectory of moving targets in real-time video data through computer vision algorithms, and to predict the future motion trajectory of moving targets through prediction algorithms.

[0070] The execution module includes an air curtain machine, which includes an adjustable speed fan unit and an adjustable angle air guide plate unit;

[0071] The environmental sensing module is used to acquire information on the indoor and outdoor temperature difference in the area covered by the air curtain machine;

[0072] The control module is used to generate control commands for controlling the fan unit and the air guide plate unit based on the real-time motion trajectory and the future motion trajectory.

[0073] In this embodiment, the present invention uses the image acquisition device of the sensing module to monitor the coverage area of ​​the air curtain machine in real time, and uses the computer vision algorithm of the data processing module to identify the real-time trajectory of moving targets such as pedestrians, thereby predicting their movement path in the next 1-3 seconds; at the same time, the environmental sensing module acquires indoor and outdoor temperature difference information. The control module integrates this information to generate precise control commands, driving the air curtain machine in the execution module to dynamically adjust the speed of the fan unit and the angle of the air guide plate unit. This achieves a leap from "passive reaction" to "active prediction", which can form an airflow barrier that accurately matches the timing and location of pedestrians in advance. This not only significantly improves the comfort experience of people without them noticing and avoids the sudden wind changes of traditional methods, but also dynamically optimizes the operating state according to the gaps in pedestrian flow and temperature differences, minimizing ineffective energy consumption and achieving a balance between high efficiency, energy saving and comfort.

[0074] Preferably, when the data processing module identifies the real-time motion trajectory of a moving target in real-time video data using computer vision algorithms, it includes the following steps:

[0075] S11: Extract the current frame image from the real-time video stream acquired by the image acquisition device at a preset frame rate;

[0076] S12: Input the current frame image into the pre-trained target detection neural network model to obtain the initial detection result set of all moving targets in the frame;

[0077] S13: Associate the detection results of the current frame with the established set of tracked targets and assign an identity ID to each detected target;

[0078] S14: For a successfully associated tracking target, based on its current observations and existing motion history, a state estimation algorithm is used to update its motion state vector, which contains position and velocity information.

[0079] S15: Based on the updated motion state vector, generate trajectory information for each moving target at timestamp t.

[0080] The object detection neural network model is a single-stage detector based on anchor boxes. Its training loss function is a weighted sum of classification loss, bounding box regression loss, and object confidence loss, as shown in the formula:

[0081] ;

[0082] in, For classifying losses, For bounding box regression loss, The target confidence loss uses the cross-entropy loss function for classification, the CIoU loss function for bounding box regression, and the binary cross-entropy loss function for target confidence loss. , and The balancing weights for each loss term.

[0083] Specifically, when associating the detection results of the current frame with the established set of tracked targets, an association strategy based on intersection-union ratio (IU) and appearance features is adopted, including:

[0084] Calculate the cost matrix between the set of detection results and the predicted set of tracked targets;

[0085] The Hungarian algorithm is used to perform optimal matching of the cost matrix, assigning to detection results with confidence scores above a threshold, and initializing new tracking trajectories for successful detections that do not match.

[0086] The state estimation algorithm employs the Kalman filter algorithm.

[0087] In this embodiment, the invention first extracts image frames from the video stream and inputs them into a single-stage detection neural network based on anchor boxes. This network is optimized through a weighted loss function that fuses classification loss, CIoU bounding box regression loss, and target confidence loss, ensuring the accuracy and robustness of detection. Subsequently, using an association strategy combining intersection-union ratio (IU) and appearance features with the Hungarian algorithm, the detection results are stably matched with existing tracked targets, and trajectories are initialized for new targets. Finally, a Kalman filter algorithm is used to estimate the state of successfully associated targets, outputting a motion state vector containing position and velocity in real time. Through layer-by-layer optimized algorithm combinations, this invention significantly improves the real-time performance, accuracy, and stability of moving target trajectory recognition in complex scenes, providing a solid and reliable perceptual data foundation for subsequent trajectory prediction and precise pre-control of air curtain machines.

[0088] Preferably, when the data processing module predicts the future trajectory of a moving target using a prediction algorithm, it includes the following steps:

[0089] S21: Trajectory sequence preprocessing, to obtain the smoothed state sequence of the moving target in the most recent T historical time steps, i.e., the historical trajectory sequence;

[0090] S22: Trajectory feature encoding, which inputs the historical trajectory sequence into the encoder network, extracts its implicit spatiotemporal features, and outputs a feature vector of fixed dimensions;

[0091] S23: Future trajectory decoding: The feature vector is input into the decoder network, and the decoder network recursively predicts the future target state to generate a set of predicted trajectories;

[0092] S24: Optimal trajectory selection: Select the optimal predicted trajectory from the predicted trajectory set according to preset rules.

[0093] The encoder and decoder together constitute a sequence-to-sequence model, the core of which is a long short-term memory network. The update formula for the encoder's long short-term memory network is as follows:

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] in, For the current time step, Let be the input vector at time step t. For the hidden state at time step t, For the cell state at time step t, For input gate, For the Gate of Oblivion For output gate, This is the hidden state of the previous time step t-1. , , and These are the weight matrices multiplied with the input, corresponding to the input gate, forget gate, output gate, and candidate cell state, respectively. , , and These are the weight matrices corresponding to the input gate, forget gate, output gate, and candidate cell state, respectively, multiplied by the previous hidden state. , , and These correspond to the bias vectors of the input gate, forget gate, output gate, and candidate cell state, respectively. It is the Sigmoid activation function. The hyperbolic tangent activation function is used. This indicates element-wise multiplication.

[0100] Among them, the decoder long short-term memory network is used as Starting with the initial state, we recursively make predictions: at each step, we take the predicted state from the previous step as input, output the predicted state for the current step, and generate the future trajectory.

[0101] In generating the predicted trajectory set, a conditional variational autoencoder or generative adversarial network structure is used, with the goal of learning the multimodal distribution present in the training data.

[0102] In this embodiment, the present invention first performs smoothing preprocessing on the historical trajectory sequence of the target, and then inputs it into an encoder-decoder model built with a long short-term memory network as its core. This model effectively captures the long-term spatiotemporal dependencies of the trajectory through its finely designed gating mechanism, providing a solid foundation for prediction. Crucially, the system innovatively solves the multimodal problem in trajectory prediction by integrating a conditional variational autoencoder or generative adversarial network structure. It can simultaneously generate multiple reasonable trajectories reflecting different future possibilities, and then select the most probable path based on optimality rules. This significantly improves the accuracy and reliability of predicting pedestrian movement intentions in complex scenarios, providing crucial forward-looking information for the air curtain machine to achieve precise pre-control of "airflow first," fundamentally reducing the impact of system latency, and is a core guarantee for achieving a seamless and comfortable experience and high energy efficiency.

[0103] Preferably, the data processing module also includes, during operation, coarse-grained attribute analysis of the moving target that does not involve personal identification, specifically including:

[0104] A11: Infer the thickness of clothing based on the target's contour information;

[0105] A12: Identify whether a moving target is carrying large items;

[0106] A13: Based on the height and gait characteristics of the movement target, they are classified into the age groups of children, adults, and the elderly.

[0107] In this embodiment, the present invention further performs rigorous coarse-grained attribute analysis on moving targets based on anonymized data. By parsing the target contour information, it infers clothing thickness, detects whether the target is carrying large items, and classifies them into age groups (children / adults / elderly) based on height and gait characteristics. This solution, while absolutely protecting personal privacy, enables the system to perceive and understand the potential needs of different pedestrians, thereby intelligently adjusting airflow strategies (such as reducing wind intensity for those wearing heavy clothing, enhancing airflow shielding for those carrying large items, and switching to a gentler mode for children or the elderly). Ultimately, it achieves an upgrade from "treating everyone the same" to "being considerate," significantly improving the personalized comfort experience for users when passing through the air curtain.

[0108] Preferably, when the data processing module infers the clothing thickness based on the target's contour information, it specifically includes:

[0109] S31: Obtain the image region and trajectory information of the moving target. The image region is defined by the bounding box provided by the target detection and tracking module, and the trajectory information includes the position sequence in consecutive frames.

[0110] S32: Extract the binary contour mask of the moving target from the image region;

[0111] S33: Calculate the aspect ratio and pixel area of ​​the contour mask;

[0112] S34: The aspect ratio and pixel area are fused with the current outdoor temperature value obtained from the environmental sensor to form a feature vector;

[0113] S35: Input the fused feature vector into the classifier, and output the clothing classification label representing the thickness of the clothing through the classifier. The clothing classification label includes light and thin and thick.

[0114] In this embodiment, the present invention employs a multimodal data fusion intelligent method to infer clothing thickness: First, a binary contour mask of the moving target is extracted from the image, and its aspect ratio and pixel area are calculated as morphological features. Then, these visual features are fused with real-time outdoor temperature data to form a comprehensive feature vector, which is finally input into a classifier to obtain a "light" or "heavy" clothing label. This solution cleverly combines visual morphological information with the environmental temperature context, making the judgment more accurate and logical. This allows the air curtain machine to intelligently adjust the airflow according to the actual physical needs of pedestrians (such as appropriately reducing the wind speed when wearing heavy clothing in winter), which not only significantly improves comfort but also avoids energy waste caused by excessive airflow, achieving truly intelligent and humanized environmental control.

[0115] Preferably, when performing attribute analysis that does not involve personal identification in parallel based on image region and trajectory information, the method also includes inferring the age group attribute of the moving target, specifically:

[0116] S41: Calculate the average speed and gait period characteristics of a moving target based on trajectory information;

[0117] S42: Based on the image region, estimate the pixel height of the moving target, and combine the camera intrinsic parameters to estimate the absolute height of the moving target;

[0118] S43: Input the estimated absolute height, average speed and gait features into the classifier, and output the age classification label representing the age group. The age classification label includes children, adults and the elderly.

[0119] Among them, targets with estimated height significantly below the threshold and small stride and slow speed tend to be classified as children; targets with normal height but slow speed and gait characteristics consistent with the elderly tend to be classified as elderly.

[0120] In this embodiment, the present invention employs a multi-feature fusion decision-making method when inferring age group attributes: by analyzing the absolute height, average walking speed, and gait cycle characteristics (such as stride length and rhythm) of the moving target, a classifier comprehensively determines whether the target belongs to a child, adult, or elderly person. This scheme requires no personal identification information; it achieves highly reliable classification solely based on anonymized data through a combination of natural behavioral characteristics such as height (e.g., significantly short stature), speed (e.g., fast cadence but slow pace), and gait (e.g., unsteady gait, short stride). Its beneficial effect lies in enabling the system to proactively identify and care for vulnerable groups (e.g., switching to a gentler, safer airflow mode for children and the elderly), significantly improving the seamless experience for users of different age groups, and demonstrating the humanistic care and precise service capabilities of the intelligent system.

[0121] In addition to performing attribute analysis in parallel based on image region and trajectory information without involving personal identification, the method also includes identifying whether a moving target is carrying a large object, specifically:

[0122] The image region containing the moving target is input into a pre-trained target detection neural network, which is specifically designed to detect large items such as suitcases, trolleys, and large packages.

[0123] If the above-mentioned large item is detected, and the overlap between the bounding box of the item and the bounding box of the moving target exceeds a preset threshold, it is determined that the moving target is carrying a large item, and the category and size information of the item are recorded.

[0124] In this embodiment, the present invention employs a precise discrimination process when identifying large items: the image region containing the moving target is input into a specially trained target detection neural network to identify specific large item categories such as suitcases and trolleys; if such an item is detected and its bounding box overlaps with the bounding box of the moving target exceeding a threshold, it is determined to be a carrying relationship. This scheme can accurately sense whether pedestrians are carrying large items and their dimensions, thereby enabling the system to proactively trigger an enhanced operating mode, generating a stronger and wider-coverage airflow barrier to ensure effective isolation of large intruders. This significantly improves the reliability of environmental isolation in practical scenarios such as logistics handling and baggage check-in, while avoiding energy waste caused by continuous high-intensity operation.

[0125] Preferably, the data processing module is also used to identify preset behaviors based on the real-time motion trajectory of the moving target. The preset behaviors include staying, lingering, and suddenly turning back in the area covered by the air curtain machine.

[0126] When lingering and loitering behavior is detected, the control module generates instructions to reduce the wind speed in the corresponding area and switch to a gentle mode.

[0127] When a sudden reversal is detected, the control module generates a command to cancel the pre-loaded acceleration control of the corresponding moving target.

[0128] In this embodiment, the present invention further intelligently identifies specific behaviors (such as stopping, lingering, and sudden reversal) by analyzing the real-time trajectory of moving targets. When a person is detected stopping or lingering at a doorway, the system automatically reduces the wind speed in the corresponding area and switches to a gentle mode to avoid discomfort caused by continuous strong winds. When a pedestrian is detected to suddenly turn back, the system immediately cancels the prepared acceleration command. This enables the system to understand and respond to complex human behaviors, effectively avoiding energy waste caused by people temporarily changing their intentions, while significantly improving the humanized interactive experience in real-world scenarios, demonstrating a key advancement in intelligent systems from mechanical response to proactive adaptation.

[0129] Preferably, the control module, when in operation, specifically includes:

[0130] S51: Continuously receives real-time motion trajectory of moving targets, predicted future motion trajectory and attribute analysis results from the data processing module, and obtains indoor and outdoor temperature difference data from the environmental perception module;

[0131] S52: Determine the current operating mode that the system should enter. Operating modes include standby mode, single-target pre-action mode, multi-target collaborative mode, and special behavior mode.

[0132] S53: Generate control commands containing the target wind speed value and the target wind deflector angle based on the determined working mode;

[0133] S54: Send control commands to the execution module; for commands based on predicted trajectories, perform timing calculations to achieve early triggering.

[0134] In this embodiment, the core advantage of the control module of the present invention lies in its highly intelligent decision-making process: by continuously integrating multi-source information such as real-time / predicted trajectories, attribute analysis results, and environmental temperature differences from the data processing module, it dynamically judges and switches to the most suitable working mode (such as standby, single-target pre-action, multi-target collaboration, or special behavior mode), thereby generating precise wind speed and direction control commands and triggering them in advance based on predictive information. This achieves a leap from simple feedback to forward-looking decision-making. Through multi-mode collaboration and pre-action mechanisms, it ensures the high accuracy and adaptability of airflow control, thereby maximizing energy efficiency and personnel comfort while meeting the high isolation requirements in different scenarios, achieving the optimal balance between intelligence and energy saving.

[0135] Preferably, when the control module determines that the current system has entered the single-target pre-action mode, the specific workflow includes:

[0136] S61: When it is predicted that a single moving target will enter the air curtain machine coverage area within a preset time threshold, calculate its expected arrival location and time based on the predicted trajectory.

[0137] S62: Based on the analysis results of the moving target's attributes and the indoor-outdoor temperature difference, fine-tune the target wind speed and the angle of the wind deflector.

[0138] S63: Based on the estimated arrival time, generate an acceleration command one system delay time in advance to ensure that the airflow is stably established before the moving target arrives.

[0139] In this embodiment, the core process of the present invention in the single-target pre-action mode is as follows: when a single target is predicted to enter the air curtain zone within a set time, the system immediately calculates the precise arrival time and location based on its predicted trajectory, and fine-tunes the wind speed and angle parameters in conjunction with its attributes (such as age and clothing) and ambient temperature difference, ultimately issuing a control command one system delay cycle in advance. Through precise spatiotemporal prediction and personalized parameter adjustment, the airflow barrier is ensured to always be established stably half a step ahead of the target, fundamentally eliminating the response delay problem of traditional air curtain machines, achieving a seamless and comfortable passage experience where "the wind arrives before the person does," while avoiding energy waste. This is a key innovation for improving experience and energy efficiency.

[0140] Preferably, when the control module determines that the current system has entered the multi-target pre-action mode, the specific workflow includes:

[0141] S71: When there are multiple predicted trajectories of moving targets at the same time and the trajectories intersect within the coverage area of ​​the air curtain machine, calculate the smallest continuous area that can cover all relevant trajectories.

[0142] S72: The generated control commands are used to control the fan unit to rise to a wind speed that can cover the entire area and to control the air guide plate unit to swing.

[0143] In this embodiment, under a multi-target collaborative mode, the present invention intelligently generates control commands by calculating the minimum coverage area of ​​the predicted trajectories of multiple moving targets. This enables the fan unit to be increased to a uniform wind speed that effectively covers the entire area, and drives the air guide plate unit to oscillate periodically. This successfully solves the problem of control lag or over-response in traditional air curtain machines when facing groups of pedestrians. It achieves efficient and comprehensive isolation of continuous crowds with a single, continuous "airflow sweep" rather than chaotic multiple airflows, significantly improving environmental isolation efficiency and passage comfort in densely populated scenarios, while avoiding the huge energy waste caused by independently delivering full-speed air to each target.

[0144] Preferably, when the control module determines that the current system has entered a special behavior mode, it specifically includes:

[0145] A21: Dwelling Response Sub-mode. When a moving target is detected to have stayed in the area covered by the air curtain machine for more than a time threshold, an instruction is generated to reduce the wind speed in the local area where the target is located.

[0146] A22: Turnback Cancellation Sub-mode. When a moving target is detected to suddenly turn back, an instruction is generated to cancel the target's pre-loaded acceleration instruction and restore the wind speed control parameters to the state before the turnback.

[0147] In this embodiment, the present invention possesses sophisticated scenario-response capabilities under specific behavioral modes: when the system detects that a person has lingered in the air curtain area for more than a threshold, it immediately activates a lingering response sub-mode, automatically reducing the wind speed in the corresponding local area to avoid causing discomfort; when a pedestrian is detected suddenly turning back, a turn-back cancellation sub-mode is triggered, immediately canceling the pre-loaded acceleration command and restoring the original operating state. This mechanism enables the system to intelligently respond to real-world scenarios such as talking at the door, waiting for someone, or temporarily changing one's mind, effectively eliminating ineffective energy consumption. It also demonstrates the system's deep understanding and high adaptability to human behavior, further optimizing energy efficiency and user-friendly experience.

[0148] Preferably, when the control module generates control commands containing the target wind speed value and the target wind deflector angle according to the determined working mode, it also includes a wind direction compensation energy-saving strategy based on the indoor and outdoor temperature difference, specifically:

[0149] When no moving target is detected, the control fan unit operates at a lower base wind speed, and the control air guide plate unit tilts towards the outdoor side at a compensation angle to achieve better airtightness isolation with low energy consumption.

[0150] When a moving target is detected, the air guide plate unit is controlled to return to the optimal isolation angle.

[0151] In this embodiment, the present invention further introduces an energy-saving strategy based on temperature difference and wind direction compensation: when no one is passing by, the system controls the fan to operate at a lower base wind speed, while tilting the air guide plate towards the outdoor side at a specific angle to form a tilted, more "enveloping" airflow barrier. This significantly reduces energy consumption while enhancing the insulation effect against hot and cold air through angle compensation. Once a pedestrian is detected, the air guide plate immediately returns to the optimal isolation angle. This solution intelligently decouples wind speed and wind direction control, achieving the best balance between energy saving and basic insulation / insulation performance by optimizing airflow direction during idle periods rather than simply sacrificing airflow volume, thus significantly improving the overall energy efficiency of the system.

[0152] Example 1: Application at the entrance of a high-end commercial supermarket

[0153] This embodiment demonstrates the application of the system at the entrance of a commercial supermarket with high foot traffic and high requirements for customer experience.

[0154] The intelligent air curtain system of this invention is installed above the glass doors of the supermarket. The sensing module consists of two high-definition cameras installed on the top of the doorway, covering an area of ​​approximately five meters inside and outside the door. The temperature and humidity sensors of the environmental sensing module are installed in the indoor and outdoor shaded areas, respectively.

[0155] On a winter evening, the outdoor temperature was 5°C and the indoor temperature was 22°C. The system was in low-power standby mode, with the fan running at a low base wind speed. At the same time, the air guide plate was tilted 45 degrees towards the outside, using a wind direction compensation strategy to maintain basic insulation while saving energy.

[0156] At this moment, an adult wearing a heavy down jacket (Target A), carrying shopping bags, is steadily walking towards the door. The camera in the data processing module captures the image, and the target detection and tracking algorithm quickly identifies Target A and assigns it a unique ID. Through Kalman filtering, the system stably tracks its trajectory. Based on its movement history over the past second, the trajectory prediction module predicts that it will pass through the right side of the door in 1.5 seconds using a Seq2Seq model. Simultaneously, the attribute analysis module works in parallel: based on its broad silhouette and the low outdoor temperature, it determines that it is "heavily dressed"; no large items are detected; and based on its height and gait, it is determined to be an "adult".

[0157] After receiving this information, the control module determines to enter the "single-target pre-action mode." Based on the predicted trajectory and arrival time, and considering the "adult" attribute and the large indoor-outdoor temperature difference, it calculates that the optimal wind speed should be medium-high (one level lower if it is for a child). 0.5 seconds before target A actually steps into the doorway, the control module issues a command in advance, the fan unit smoothly accelerates to the predetermined wind speed, and the air deflector simultaneously rotates to 90 degrees vertically downwards. When target A passes through the air curtain, the airflow barrier has been stably established; he only feels a gentle and effective warm breeze and is not impacted by strong winds. After passing through the doorway, the system predicts that his trajectory has entered the room and returns to standby mode after 1 second.

[0158] Immediately afterward, the family of three (Target B: father, Target C: mother, Target D: child) walked side by side towards the door, chatting and laughing. The system immediately switched to "multi-target collaborative mode." It predicted three almost parallel trajectories and calculated a minimum continuous area that could cover all three people. Subsequently, the system generated instructions to raise the fan to a higher, uniform wind speed and control the air deflectors to automatically oscillate slightly, generating a wide, sweeping airflow that simultaneously covered all three, effectively preventing cold air from entering. The attribute analysis module specifically identified Target D as the "child," so the system fine-tuned the airflow intensity while ensuring the overall air curtain remained effective, avoiding discomfort to the child.

[0159] When no one is passing by, the system continuously runs a wind direction compensation energy-saving strategy, significantly reducing energy consumption during off-peak hours. On one occasion, a customer stopped at the door to answer a phone call. After 3 seconds, the behavior recognition module determined it to be a "stopping" behavior, and the system immediately activated the "stopping response sub-mode," reducing the wind speed in the customer's area to prevent strong winds from blowing for an extended period, demonstrating human-centered care.

[0160] Example 2: Application of Cold Chain Logistics Warehouse Loading and Unloading Platform

[0161] This embodiment demonstrates the application of the system in an industrial environment with extreme requirements for energy consumption and isolation efficiency—a cold chain logistics warehouse loading and unloading platform.

[0162] This system was installed at the giant lifting door of the warehouse. The environmental sensing module detected an outdoor temperature of 32℃, while the temperature inside the cold storage was -20℃, a huge temperature difference. This placed extremely high demands on energy consumption control.

[0163] Normally, the warehouse door is closed and the system is in standby mode. When the lifting door opens and a forklift is in operation, the system activates. At this time, a forklift fully loaded with goods (target E) is driving from outside into the warehouse. The system not only identifies the forklift driver, but more importantly, its object detection neural network identifies the forklift itself and the large volume of goods it carries as a single "large item." The trajectory prediction module accurately predicts the forklift's arrival time and path at the door based on its relatively high travel speed. The attribute analysis results include "carrying a large item" and "moving at high speed."

[0164] The control module then enters an enhanced "single-target pre-action mode." It instructs the fan unit to operate at its maximum speed and adjusts the air deflectors to the optimal angle to ensure effective sealing of the large intrusion cross-section created by the forklift and goods, preventing significant loss of cold air. The entire process involves precise prediction, with airflow almost synchronized with the forklift, achieving highly efficient isolation.

[0165] Subsequently, two forklifts, one moving out and the other in (targets F and G), approached the door almost simultaneously. The system entered "multi-target collaborative mode." It calculated the minimum coverage area of ​​the two opposing trajectories and generated commands to control the fans to operate at maximum speed. At the same time, it could dynamically adjust the angle of the air guide vanes to form a powerful and continuous airflow wall at the door, simultaneously handling traffic flow from both directions. This ensured that the cold storage's airtightness was maximized while operating efficiently.

[0166] During a break in operations, two warehouse employees (targets H and I) briefly exchanged words inside the doorway. The behavior recognition module quickly identified this "pausing" behavior, and the system immediately triggered the pause response sub-mode in the "Special Behavior Mode," automatically reducing the wind speed in that area to a comfortable low level to prevent the employees from being exposed to strong cold air. Once they finished their conversation and left, the system immediately resumed regular monitoring. Furthermore, when the system predicts that a forklift will suddenly reverse (turnaround behavior) just before entering the door, it will immediately cancel the pre-loaded acceleration command to avoid energy waste.

[0167] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A computer vision-based intelligent airflow control system for air curtain machines, characterized in that: include: The sensing module includes an image acquisition device located near the area covered by the air curtain machine, used to acquire real-time video data containing moving targets; The data processing module is used to identify the real-time motion trajectory of moving targets in real-time video data through computer vision algorithms, and to predict the future motion trajectory of moving targets through prediction algorithms. The execution module includes an air curtain machine, which includes an adjustable speed fan unit and an adjustable angle air guide plate unit; The environmental sensing module is used to acquire information on the indoor and outdoor temperature difference in the area covered by the air curtain machine; The control module is used to generate control commands for controlling the fan unit and the air guide plate unit based on the real-time motion trajectory and the future motion trajectory.

2. The intelligent airflow control system for an air curtain machine based on computer vision according to claim 1, characterized in that: When the data processing module identifies the real-time motion trajectory of a moving target in real-time video data using computer vision algorithms, it includes the following steps: S11: Extract the current frame image from the real-time video stream acquired by the image acquisition device at a preset frame rate; S12: Input the current frame image into the pre-trained target detection neural network model to obtain the initial detection result set of all moving targets in the frame; S13: Associate the detection results of the current frame with the established set of tracked targets and assign an identity ID to each detected target; S14: For a successfully associated tracking target, based on its current observations and existing motion history, a state estimation algorithm is used to update its motion state vector, which contains position and velocity information. S15: Based on the updated motion state vector, generate trajectory information for each moving target at timestamp t.

3. The intelligent airflow control system for an air curtain machine based on computer vision according to claim 2, characterized in that: When the data processing module predicts the future trajectory of a moving target using a prediction algorithm, it includes the following steps: S21: Trajectory sequence preprocessing, to obtain the smoothed state sequence of the moving target in the most recent T historical time steps, i.e., the historical trajectory sequence; S22: Trajectory feature encoding, which inputs the historical trajectory sequence into the encoder network, extracts its implicit spatiotemporal features, and outputs a feature vector of fixed dimensions; S23: Future trajectory decoding: The feature vector is input into the decoder network, and the decoder network recursively predicts the future target state to generate a set of predicted trajectories; S24: Optimal trajectory selection: Select the optimal predicted trajectory from the predicted trajectory set according to preset rules.

4. The intelligent airflow control system for an air curtain machine based on computer vision according to claim 3, characterized in that: The data processing module, during operation, also includes coarse-grained attribute analysis of moving targets that does not involve personal identification, specifically including: A11: Infer the thickness of clothing based on the target's contour information; A12: Identify whether a moving target is carrying large items; A13: Based on the height and gait characteristics of the movement target, they are classified into the age groups of children, adults, and the elderly.

5. The intelligent airflow control system for an air curtain machine based on computer vision according to claim 4, characterized in that: When the data processing module infers the clothing thickness based on the target's contour information, it specifically includes: S31: Obtain the image region and trajectory information of the moving target. The image region is defined by the bounding box provided by the target detection and tracking module, and the trajectory information includes the position sequence in consecutive frames. S32: Extract the binary contour mask of the moving target from the image region; S33: Calculate the aspect ratio and pixel area of ​​the contour mask; S34: The aspect ratio and pixel area are fused with the current outdoor temperature value obtained from the environmental sensor to form a feature vector; S35: Input the fused feature vector into the classifier, and output the clothing classification label representing the thickness of the clothing through the classifier. The clothing classification label includes light and thin and thick.

6. The intelligent airflow control system for an air curtain machine based on computer vision according to claim 5, characterized in that: When the control module is working, it specifically includes: S51: Continuously receives real-time motion trajectory of moving targets, predicted future motion trajectory and attribute analysis results from the data processing module, and obtains indoor and outdoor temperature difference data from the environmental perception module; S52: Determine the current operating mode that the system should enter. Operating modes include standby mode, single-target pre-action mode, multi-target cooperative mode, and special behavior mode. S53: Generate control commands containing the target wind speed value and the target wind deflector angle based on the determined working mode; S54: Send control commands to the execution module; for commands based on predicted trajectories, perform timing calculations to achieve early triggering.

7. The intelligent airflow control system for an air curtain machine based on computer vision according to claim 6, characterized in that: When the control module determines that the current system has entered the single-target pre-action mode, the specific workflow includes: S61: When it is predicted that a single moving target will enter the air curtain machine coverage area within a preset time threshold, calculate its expected arrival location and time based on the predicted trajectory; S62: Based on the analysis results of the moving target's attributes and the indoor-outdoor temperature difference, fine-tune the target wind speed and the angle of the wind deflector. S63: Based on the estimated arrival time, generate an acceleration command one system delay time in advance to ensure that the airflow is stably established before the moving target arrives.

8. The intelligent airflow control system for an air curtain machine based on computer vision according to claim 7, characterized in that: When the control module determines that the current system has entered the multi-target pre-action mode, the specific workflow includes: S71: When there are multiple predicted trajectories of moving targets at the same time and the trajectories intersect within the coverage area of ​​the air curtain machine, calculate the smallest continuous area that can cover all relevant trajectories. S72: The generated control commands are used to control the fan unit to increase the wind speed to cover the entire area and to control the air guide plate unit to swing.

9. The intelligent airflow control system for an air curtain machine based on computer vision according to claim 8, characterized in that: When the control module determines that the current system has entered a special behavior mode, it specifically includes: A21: Dwelling Response Sub-mode. When a moving target is detected to have stayed in the area covered by the air curtain machine for more than a time threshold, an instruction is generated to reduce the wind speed in the local area where the target is located. A22: Turnback Cancellation Sub-mode. When a moving target is detected to suddenly turn back, an instruction is generated to cancel the target's pre-loaded acceleration instruction and restore the wind speed control parameters to the state before the turnback.

10. The intelligent airflow control system for an air curtain machine based on computer vision according to claim 9, characterized in that: When the control module generates control commands containing the target wind speed value and the target wind deflector angle according to the determined working mode, it also includes a wind direction compensation energy-saving strategy based on the indoor and outdoor temperature difference, specifically: When no moving target is detected, the control fan unit operates at a lower base wind speed, and the control air guide plate unit tilts towards the outdoor side at a compensation angle to achieve better airtightness isolation with low energy consumption. When a moving target is detected, the air guide plate unit is controlled to return to the optimal isolation angle.

Images