A home target detection method and system

Abstract

The application discloses a kind of home target detection method and system, obtain home scene parameter and calibration, collect the environmental signal from home scene, the target signal is obtained by pre-processing the environmental signal;Wake up confidence is calculated based on the target signal, when the wake-up confidence meets the condition, start image acquisition mode and collect target image, obtain target image pair, after completion of acquisition, control the image acquisition mode stops work;The target image pair is sequentially pre-processed, registered and fused, the fusion result is input into fusion network to extract depth information, the depth information is input into the constructed determination model to obtain the determination result of distinguishing real human target or pseudo target;Similarity is calculated based on the determination result to determine target effectiveness and classification;Significantly reduce the risk of rural night no-light scene single-mode detection failure, effectively improve the detection rate of weak and small shielding target, reduce the risk of privacy leakage and terminal power consumption.

Description

Technical Field

[0001] This invention relates to the field of target detection, and more specifically to a method and system for home target detection. Background Technology

[0002] Compared to standardized urban homes, rural living environments are inherently characterized by multiple interferences, drastic changes in lighting throughout the day, weak maintenance, and strict privacy constraints. Furthermore, the focus of monitoring is on left-behind elderly, left-behind children, and vulnerable groups such as the sick and disabled, creating unique technical challenges for dual-modal target detection—a core bottleneck that traditional single-modal intelligent monitoring technologies cannot overcome. From the perspective of environmental and modal adaptation difficulties, rural homes experience significant visible light interference during the day due to haphazardly piled-up items, cluttered background textures, and frequent light and shadow fluctuations. At night, with no auxiliary lighting and complete darkness, the visible light modality becomes completely ineffective, relying solely on infrared sensing. However, temperature noise from indoor appliances, walls, and floors easily masks subtle human body temperature characteristics. In cloudy or shadowy conditions, visible light contrast is extremely low, details are lost, and infrared temperature gradients are blurred, resulting in feature degradation in both modalities. From the perspective of the characteristics and difficulties of the monitored targets, the elderly left behind in rural areas are often hunched over, sit or lie still for long periods, and have weak limb movements, resulting in incomplete visible light outlines and flat, indistinct infrared temperature characteristics; the children left behind in rural areas are small in size and easily obscured by clutter, resulting in a very small proportion of dual-modal targets and fragmented characteristics; the sick and disabled have fixed postures and remain still for long periods without dynamic changes, making it impossible to distinguish static human bodies from environmental pseudo-targets using single-modal methods. In summary, under the all-weather conditions of rural homes, both single visible light and single infrared modal methods have perception blind spots, and dual-modal complementary fusion is necessary to effectively detect vulnerable targets. Currently, there is no detection method in the industry that is adapted to the drastic changes in lighting conditions in rural homes throughout the day, addresses the needs of the elderly, children, sick, disabled, and small static targets that are obscured, utilizes a lightweight infrared-visible light dual-modal deep fusion method based on D-FINE, and has a built-in native privacy destruction mechanism. Summary of the Invention

[0003] This invention proposes a method for home target detection, comprising: S1. Obtain and calibrate home scene parameters, collect environmental signals from the home scene, and preprocess the environmental signals to obtain the target signal. S2. Calculate wake-up confidence based on the target signal. When the wake-up confidence meets the condition, start the image acquisition mode to acquire the target image and obtain the target image pair. After the acquisition is completed, control the image acquisition mode to stop working. S3. The target image pairs are preprocessed, registered and fused in sequence. The fusion result is input into the fusion network to extract depth information. The depth information is input into the construction of a judgment model to obtain a judgment result that distinguishes between real human targets and pseudo targets. S4. Calculate the similarity based on the judgment result to determine the effectiveness of the target and classify it.

[0004] Furthermore, the collection of environmental signals from the home environment specifically includes: a first channel for capturing precursor signals of limb shaking, movement, and abnormal bodily movements, and a second channel for listening to distress commands such as groans and calls for help.

[0005] Furthermore, the acquisition and calibration of home scene parameters specifically includes: initializing the low light brightness judgment range, background noise suppression coefficient, human body contour feature threshold, target identification threshold, interference filtering weight, effectiveness grading threshold, encryption key, and delay parameters.

[0006] Furthermore, the preprocessing of the environmental signal specifically includes: a time-domain differential filtering algorithm filters out irregular environmental disturbance data by comparing the amplitude change rate of adjacent time-series signals, and retains real human anomalies and voice signals.

[0007] Furthermore, the preprocessing, registration, and fusion of the target image pairs in sequence specifically includes: an adaptive thresholding algorithm that maximizes the threshold segmentation of the image foreground and background by calculating the inter-class variance, suppressing background texture, and stretching the brightness of low-light areas to restore human body contour details.

[0008] Furthermore, the preprocessing, registration, and fusion of the target image pairs in sequence specifically includes: the temperature threshold filtering model compares the pixel temperature value with the environmental thermal clutter baseline to remove static temperature pseudo-targets that are below the normal human body temperature range and enhances the local temperature gradient features.

[0009] Furthermore, after calculating the similarity based on the judgment result to determine the effectiveness of the target and classify it, the process further includes: clearing the image, encrypting the result text using an encryption algorithm before transmission, entering a low-power state, and updating the model parameters.

[0010] Furthermore, the step of inputting the depth information into the judgment model to obtain a judgment result that distinguishes between real human targets and pseudo targets specifically includes: the judgment model integrates feature response scores, boundary integrity coefficients, human body proportion constraints, and temperature range rationality parameters in multiple dimensions, and when any dimension deviates from the specific threshold, it is judged as a pseudo target.

[0011] Furthermore, this application also proposes a non-volatile storage medium on which a computer program is stored, which, when executed by a processor, implements the home target detection method as described above.

[0012] Furthermore, this application also proposes a home target detection system, comprising: The initialization monitoring module is used to acquire and calibrate home scene parameters, collect environmental signals from the home scene, and preprocess the environmental signals to obtain the target signal. The wake-up acquisition module is used to calculate the wake-up confidence level based on the target signal. When the wake-up confidence level meets the conditions, the image acquisition mode is started to acquire the target image and obtain the target image pair. After the acquisition is completed, the image acquisition mode is controlled to stop working. The feature discrimination module is used to perform preprocessing, registration and fusion of target image pairs in sequence, input the fusion result into the fusion network to extract depth information, and input the depth information into the judgment model to obtain the judgment result of distinguishing between real human targets and false targets. The verification closed-loop module is used to perform similarity calculations based on the judgment results to determine the effectiveness of the target and classify it.

[0013] This invention proposes a home target detection method and constructs a dual-modal fusion sensing system of infrared and visible light, which significantly alleviates the problem of single-modal detection failure in rural nighttime light-free scenes, effectively improves the detection rate of weak and occluded targets in complex backgrounds, and, with the addition of a triggered acquisition and instant clearing mechanism, fundamentally reduces the risk of privacy leakage in private home scenarios. Furthermore, through lightweight modification and a sleep closed-loop mechanism, it significantly reduces terminal power consumption and effectively adapts to the low computing power, weak network and low-frequency charging conditions in rural areas. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a home target detection method proposed in this invention. Detailed Implementation

[0015] This invention proposes a method for home target detection, comprising: S1. Obtain and calibrate home scene parameters, collect environmental signals from the home scene, and preprocess the environmental signals to obtain the target signal.

[0016] Among them, acquiring and calibrating home scene parameters aims to provide mathematical constraints for subsequent detection, collecting environmental signals from home scenes aims to adapt to complex environmental interference and the low dynamic characteristics of the target, and the target signal is the data retained after filtering out irregular environmental disturbances through preprocessing.

[0017] Specifically, initialize the weak light brightness judgment range, background noise suppression coefficient, human body contour feature threshold, target identification threshold, interference filtering weight, effectiveness classification threshold, encryption key, and delay parameter.

[0018] In a specific implementation scenario, considering the common characteristics of rural areas such as varying floor heights, house layouts, and obstructions from clutter, the system calibrates differentiated human body size ranges for the elderly, children, and disabled individuals, accurately filtering out interference from furniture, farm tools, clutter, and false targets caused by light and shadow. In sleep mode, the device maintains a dual-channel asynchronous ultra-low-power monitoring mechanism. The first channel continuously collects weak dynamic changes in indoor human body signals, capturing subtle limb movements of the elderly, slow movements of children, and subtle bodily movements of the disabled; the second channel continuously collects local voice signals, listening for groans of pain, calls for help, and other active distress commands. A time-domain differential filtering algorithm compares the amplitude change rates of adjacent time-series signals to filter out irregular environmental disturbances such as outdoor wind noise, poultry calls, neighborhood noise, and flickering light and shadow from doors and windows, retaining only genuine human body movements and voice signals. The system remains in sleep standby mode, without triggering any invalid calculations.

[0019] S2. Calculate the wake-up confidence level based on the target signal. When the wake-up confidence level meets the condition, start the image acquisition mode to acquire the target image, obtaining a target image pair. After acquisition, control the image acquisition mode to stop working. The wake-up confidence level is used to quantify the wake-up confidence level. The wake-up confidence level is a value calculated from the feature parameters extracted based on the target signal, and the target image pair is image data acquired based on the image acquisition mode. When the wake-up confidence level meets the wake-up condition, it is determined as a trigger signal, and the subsequent detection process is started.

[0020] In a specific implementation scenario: Construct a dedicated wake-up reliability evaluation formula: In the formula, where, This represents the credibility of the wake-up call. Represents the human body dynamic amplitude coefficient. Represents the degree of matching of speech features. This indicates a noisy and chaotic rural environment. The localization fixed weights representing Motion The localization fixed weight of Voice, represent Localized fixed weights. Localized fixed weight configuration: Satisfying normalization constraints Set a wake-up threshold specific to rural scenarios. ,when When a valid human behavior trigger signal is detected, the subsequent detection process is initiated; when If the data is detected as environmental interference, it is immediately blocked. The visible light camera and infrared thermal imaging modules are activated simultaneously and instantaneously, acquiring only one frame of visible light grayscale image and one frame of infrared temperature image at a time. No redundant color information is acquired, no video is recorded, and no continuous frame sequence is cached. After acquisition, the dual-path hardware is immediately powered off and shut down, completely eliminating the risk of background covert acquisition, persistent snapshots, and image leakage. This achieves a privacy acquisition mechanism that triggers once, acquires two frames simultaneously, and stops immediately after acquisition.

[0021] S3. The target image pairs are preprocessed, registered, and fused sequentially. The fusion result is input into a fusion network to extract depth information. This depth information is then input into a judgment model to obtain a judgment result that distinguishes between real human targets and pseudo-targets. The fusion network is a deep network derived from a hierarchical boundary regression backbone with an added cross-modal interactive attention fusion branch. The depth information is derived from the feature parameters output by the fusion network, and the judgment result is the classification judgment data output by the judgment model. The adaptive thresholding algorithm maximizes the threshold segmentation of the image by calculating the inter-class variance, suppressing background texture, and stretching the brightness of low-light areas to restore human contour details.

[0022] In a specific implementation scenario: For visible light images, an improved Otsu adaptive thresholding algorithm is used to achieve brightness equalization, noise filtering, and suppression of cluttered background textures, restoring human contour details in low-light and shadow scenes; for infrared images, a temperature threshold filtering model is constructed to filter fixed thermal clutter from indoor appliances, walls, and floors, enhance the local temperature gradient features of weak human bodies, and eliminate static temperature pseudo-target interference. After completing pixel-level registration and alignment of the dual-modal images, a day / night adaptive modal weighting factor is designed: during the daytime strong light conditions, visible light texture weighting is biased; during the nighttime complete darkness conditions, infrared temperature weighting is biased; and during cloudy / rainy days with low light conditions, dual-modal equalization fusion is used to achieve adaptive dominance of the dominant modality in different scenarios. This study employs a self-developed lightweight dual-modal fusion D-FINETransformer network, completely abandoning traditional single-modal home target detection methods and the redundant DETR architecture. Based on the original D-FINE hierarchical boundary regression backbone, it prunes the redundant attention iteration layer and feature splicing layer of the original dual-modal fusion network, and adds a cross-modal interactive attention fusion branch, a modality adaptive weight unit, an infrared temperature clutter suppression module, a visible light texture restoration module, and a weak target dual feature enhancement unit. This achieves cross-layer bidirectional fusion and complementary enhancement of visible light shallow texture details, deep semantic features, and infrared temperature gradient and contour boundary features. Through a multi-scale attention weight allocation strategy, it strengthens weak, occluded, and static human features. Relying on D-FINE's unique boundary quality evaluation branch, iteratively optimizes the target prediction box coordinates under dual-modal fusion features, corrects the boundary offset problem of weak targets in nighttime, low-light, and occluded scenes, accurately locks human targets in cluttered backgrounds in all weather conditions, and obtains four-dimensional depth information: visible light texture semantic features, infrared temperature features, boundary confidence, and feature response intensity. Based on the four-dimensional deep information fusion dual-modal feature response score, boundary integrity coefficient, human body proportion constraint, temperature range rationality multi-dimensional parameters, a comprehensive judgment score model is constructed. Combined with localized exclusive thresholds, through a cross-modal feature mutual verification mechanism, it accurately distinguishes real human targets from furniture, farm tools, light and shadow, temperature clutter pseudo-targets, and completely solves the practical problems of day and night false detection, missed detection, and perception failure of traditional single-modal models.

[0023] S4. Calculate similarity based on the judgment results to determine the effectiveness and classify the target. Similarity is used to calculate the feature similarity of detection results across multiple consecutive frames, and classification is used to characterize the target's existence state and duration of continuous dwell. The signal-to-noise ratio of the detection network is designed to dynamically adapt to the transmission strategy.

[0024] In a specific implementation scenario: using three consecutive frames of detection results as the verification period, the temporal target feature similarity is calculated: ,in, This represents the similarity of the time-series target features. Represents the target feature score of the current frame. This represents the average target feature score within the verification period. Human targets are deemed real and valid only when the target features are stable and consistent across multiple consecutive frames, meeting steady-state conditions, thus filtering out false detections caused by instantaneous light and shadow, temperature fluctuations, and clutter. A level determination is made based on the target's presence and duration of residence: Level 1: No target, unoccupied at home; Level 2: Target appears briefly, normal activity; Level 3: Target remains statically for a long time, posing a monitoring risk. After the determination, the terminal instantly and locally destroys all original images, preprocessed images, and contour feature cache images. There is no local storage, no background backup, and no hidden retention; only text-based target detection results and valid level data are retained. An independent, lightweight XOR encryption algorithm is used. Complete data encryption, among which, This represents the encrypted target detection result text. This represents the original target detection result text to be encrypted. This represents the encryption key. Represents the bitwise XOR operator. Adapted to low-power hardware, it only encrypts and uploads minimal results such as target presence status, target credibility, and detection time, reducing data transmission volume by over 99% compared to traditional video monitoring. It monitors the signal-to-noise ratio of rural home networks in real time and dynamically adapts transmission strategies: immediate upload when the network is good; local caching, resume interrupted transmission, and automatic retransmission when the network is weak and experiencing lag; real-time monitoring of remaining device battery power, forcibly suspending non-critical data transmission tasks and prioritizing power supply to the signal monitoring channel when the battery level falls below a preset threshold. It forcibly triggers deep sleep mode for the entire machine every 0.5 seconds, shutting down the acquisition, inference, and transmission modules. A localized iterative optimization mechanism is constructed, adaptively correcting day / night mode weights, infrared thermal clutter suppression parameters, visible light weak light compensation coefficients, and weak target boundary regression thresholds with a cycle of 20 detections. ,in, Represents the updated model parameters. This represents the historical model parameters before the update. This represents the real-time model parameters collected during the current detection cycle. 0.9 and 0.1 represent the fusion weighting coefficients of historical parameters and real-time parameters, respectively.

[0025] Furthermore, this application also proposes a non-volatile storage medium. This non-volatile storage medium is used to store a computer program. Specifically, when the computer program stored therein is executed by a processor, it implements the aforementioned home target detection method. In one specific implementation scenario, the non-volatile storage medium has a read / write attribute and a generalized carrier type declaration, and when the computer program stored therein is executed by a processor, it implements the aforementioned home target detection method steps.

[0026] Furthermore, this application proposes a home target detection system. This system, an independent protected object of the system class, includes an initialization monitoring module for acquiring and calibrating home scene parameters, collecting environmental signals from the home scene, and preprocessing the environmental signals to obtain target signals; a wake-up acquisition module for calculating wake-up confidence based on the target signals, and when the wake-up confidence meets the conditions, initiating an image acquisition mode to acquire target images, obtaining target image pairs, and controlling the image acquisition mode to stop working after acquisition; a feature discrimination module for sequentially preprocessing, registering, and fusing the target image pairs, inputting the fusion result into a fusion network to extract depth information, and inputting the depth information into a judgment model to obtain a judgment result distinguishing between real human targets and pseudo-targets; and a verification closed-loop module for calculating similarity based on the judgment result to determine the effectiveness and grading of the targets. In a specific implementation scenario, the data interaction relationships between the modules strictly correspond one-to-one with the input-output closed loop of the home target detection method.

[0027] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for home target detection, characterized in that, include: S1. Obtain and calibrate home scene parameters, collect environmental signals from the home scene, and preprocess the environmental signals to obtain the target signal. S2. Calculate the wake-up confidence level based on the target signal. When the wake-up confidence level meets the condition, start the image acquisition mode to acquire the target image and obtain the target image pair. After the acquisition is completed, control the image acquisition mode to stop working. S3. The target image pairs are preprocessed, registered and fused in sequence. The fusion result is input into the fusion network to extract depth information. The depth information is input into the construction of a judgment model to obtain a judgment result that distinguishes between real human targets and false targets. S4. Calculate the similarity based on the judgment result to determine the effectiveness of the target and classify it.

2. The home target detection method as described in claim 1, characterized in that, The environmental signals collected from the home environment specifically include: a first channel for capturing precursor signals of limb shaking, movement, and abnormal bodily movements; and a second channel for listening to groans and calls for help.

3. A home target detection method as described in claim 1, characterized in that, The acquisition and calibration of home scene parameters specifically includes: initializing the low light brightness judgment range, background noise suppression coefficient, human body contour feature threshold, target identification threshold, interference filtering weight, effectiveness grading threshold, encryption key, and delay parameter.

4. A home target detection method as described in claim 1, characterized in that, The preprocessing of the environmental signal specifically includes: a time-domain differential filtering algorithm that filters out irregular environmental disturbance data by comparing the amplitude change rate of adjacent time-series signals, while retaining real human anomalies and voice signals.

5. A home target detection method as described in claim 1, characterized in that, The preprocessing, registration, and fusion of the target image pairs in sequence specifically include: an adaptive threshold algorithm that maximizes the threshold segmentation of the image foreground and background by calculating the inter-class variance, suppressing background texture, and stretching the brightness of low-light areas to restore human body contour details.

6. A home target detection method as described in claim 1, characterized in that, The preprocessing, registration, and fusion of the target image pairs in sequence specifically includes: the temperature threshold filtering model compares the pixel temperature value with the environmental thermal clutter baseline to remove static temperature pseudo-targets that are below the normal human body temperature range and enhances the local temperature gradient features.

7. A home target detection method as described in claim 1, characterized in that, The process of calculating similarity based on the judgment result to determine the effectiveness of the target and classify it further includes: clearing the image, encrypting the result text using an encryption algorithm before transmission, entering a low-power state and updating the model parameters.

8. A home target detection method as described in claim 1, characterized in that, The process of inputting the depth information into the judgment model to obtain a judgment result that distinguishes between real human targets and pseudo targets specifically includes: the judgment model integrates feature response scores, boundary integrity coefficients, human body proportion constraints, and temperature range rationality parameters. When any dimension deviates from the specified threshold, it is judged as a pseudo target.

9. A home target detection method as described in claim 1, characterized in that, It includes a non-volatile storage medium storing a computer program that, when executed by a processor, implements the home target detection method as described in any one of claims 1-8.

10. A home target detection system, characterized in that, include: The initialization monitoring module is used to acquire and calibrate home scene parameters, collect environmental signals from the home scene, and preprocess the environmental signals to obtain the target signal. The wake-up acquisition module is used to calculate the wake-up confidence level based on the target signal. When the wake-up confidence level meets the conditions, the image acquisition mode is started to acquire the target image and obtain the target image pair. After the acquisition is completed, the image acquisition mode is controlled to stop working. The feature discrimination module is used to perform preprocessing, registration and fusion of target image pairs in sequence, input the fusion result into the fusion network to extract depth information, and input the depth information into the judgment model to obtain the judgment result of distinguishing between real human targets and false targets. The verification closed-loop module is used to perform similarity calculations based on the judgment results to determine the effectiveness of the target and classify it.

Images