An AI vision recognition-based target object targeting agent monitoring method and device

Abstract

This invention discloses a target agent monitoring method and device based on AI visual recognition, belonging to the field of termite monitoring technology. It aims to solve problems such as high reliance on manual intervention, high misjudgment rate, and inefficient data transmission in traditional target agent monitoring. The device incorporates a hybrid power supply system of slow-release target agent and solar energy, supporting low-power operation. The core solution includes: fusing vibration and infrared sensor signals, triggering image acquisition through feature analysis; identifying target objects through background subtraction, multi-scale image analysis, and feature matching, dynamically optimizing detection thresholds; calculating activity intensity after video acquisition by combining target tracking and gridded analysis, optimizing transmission based on intensity and quota; and automatically adjusting supplementary lighting according to ambient light, ensuring image quality through pixel compensation and brightness adjustment. This solution constructs a closed-loop system of multimodal precise triggering and intelligent recognition, improving the intelligence and efficiency of termite monitoring.

Description

Technical Field

[0001] This invention relates to the field of termite monitoring technology, and more specifically to a method and device for monitoring target agents based on AI visual recognition. Background Technology

[0002] Visual monitoring of termites in complex environments is an important technological direction in the field of intelligent sensing. Its core challenge lies in overcoming the impact of lighting variations, background noise, and the dynamic characteristics of the target on detection accuracy. Existing visual monitoring solutions generally suffer from the following technical bottlenecks:

[0003] In the image acquisition and preprocessing stages, traditional equipment relies on fixed illumination compensation strategies. When faced with non-uniform lighting scenarios such as low light or direct strong light, problems such as blurred termite outlines and insufficient contrast easily arise, leading to the failure of subsequent feature extraction. Single-modal detection methods based on a single camera lack the ability to fuse and analyze auxiliary signals such as vibration and infrared, making it difficult to accurately separate targets in complex backgrounds. The false negative and false positive rates increase significantly with increasing environmental complexity. In target recognition and behavior analysis, existing algorithms mostly employ a single-frame independent processing mode, failing to utilize the spatiotemporal correlation information between consecutive frames to establish termite movement trajectories, making it difficult to assess the continuity and regularity of colony activities. Feature extraction for dynamic termites relies on fixed-parameter models, exhibiting poor adaptability to termites of different shapes and sizes, especially with a sharp drop in recognition accuracy in scenarios involving partial occlusion or changes in posture. In the data processing and transmission stage, no data optimization mechanism based on termite activity characteristics has been established. The transmission of redundant frames often leads to excessive bandwidth consumption, while key frames lose crucial details due to compression distortion. The real-time analysis module lacks adaptive adjustment capabilities, and fixed threshold settings cannot balance detection sensitivity and anti-interference capabilities under different monitoring scenarios, requiring frequent manual intervention and calibration. The level of intelligence urgently needs to be improved.

[0004] Existing technologies have significant shortcomings in multi-source information fusion strategies, adaptive feature extraction models, and intelligent transmission optimization mechanisms. Therefore, this invention proposes a target agent monitoring method and device based on AI visual recognition. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a target agent monitoring method and device based on AI visual recognition. By fusing vibration and infrared signals in real time for multimodal detection, target image recognition based on composite similarity, adaptive supplementary lighting adjustment, and transmission optimization driven by activity intensity, the invention solves the problems of missed detection and false detection caused by environmental interference, insufficient recognition accuracy in low light, redundant data transmission, and fixed detection thresholds in traditional monitoring devices. It achieves accurate triggering, efficient identification, and energy-saving transmission of target activity in complex environments.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A target agent monitoring method based on AI visual recognition includes:

[0008] Step S1: Real-time acquisition and fusion of device vibration signal and infrared signal, extraction of vibration frequency features, temperature gradient features and moving speed features, by configuring a judgment threshold to determine whether the vibration signal and infrared signal are both effectively triggered. If so, the detection area overlap analysis is performed based on the phase difference positioning of the vibration source and the triangulation measurement coordinates of the heat source. If there is an overlapping detection area, the camera is triggered to acquire images.

[0009] Step S2: Perform background subtraction processing on the acquired image to extract the dynamic foreground region, segment the foreground target region, and determine whether there is a target object in the acquired image by combining multi-scale image similarity matching with composite similarity calculation of shape features and texture features. If there is a target object, video acquisition is performed; and the determination threshold is adjusted based on the target object recognition result.

[0010] Step S3: After starting video acquisition, the video frames are uploaded to the monitoring platform. The monitoring platform is used to extract target features from the video frames, generate speed sequence data based on the target's movement trajectory tracking, and obtain the target's activity intensity by combining gridded distribution statistics and time window smoothing, and optimize the transmission.

[0011] Step S4: During image and video acquisition, the brightness of the supplementary light is automatically adjusted according to the ambient light. During image acquisition, pixel-level light intensity compensation is performed based on grayscale mean and contrast evaluation. During video acquisition, a smooth brightness gradient adjustment is used until the preset image quality threshold is reached.

[0012] Specifically, the steps to trigger the camera to capture images include:

[0013] The vibration sensor monitors the vibration signal of the device in real time, and the infrared sensor synchronously monitors the infrared signal of the heat source inside the device, forming a multi-dimensional detection signal.

[0014] Preprocessing of vibration and infrared signals for target objects: The vibration signal is filtered by a bandpass filter to remove environmental noise and then normalized; the infrared signal is distinguished from the background and dynamic heat source based on a Gaussian mixture model, and the temperature gradient and moving speed of the moving target are extracted.

[0015] The confidence level of the vibration signal is calculated by weighted fusion of effective vibration number and frequency characteristics, and the confidence level of the infrared signal is calculated by mapping the ratio of temperature gradient to moving speed.

[0016] Configure judgment thresholds, including vibration threshold and infrared threshold. When the confidence level of the vibration signal is greater than the vibration threshold, the vibration signal is determined to be effectively triggered. When the confidence level of the infrared signal is greater than the infrared threshold, the infrared signal is determined to be effectively triggered.

[0017] Specifically, the steps to trigger the camera to capture images also include:

[0018] When both vibration and infrared signals are effectively triggered, the detection area is divided according to the layout of the multimodal sensors. The vibration sensor determines the vibration source detection area by phase difference positioning, and the infrared sensor determines the heat source detection area by triangulation.

[0019] Determine whether the detection area effectively triggered by the vibration signal overlaps with the detection area effectively triggered by the infrared signal; if they overlap, trigger the camera to capture an image; otherwise, do not perform any processing.

[0020] Based on the identification results of whether a target object exists in the image captured by the camera, the judgment threshold is adjusted accordingly, including:

[0021] If it is determined that there is no target object in the image acquired by the trigger, the trigger determination threshold is adjusted accordingly.

[0022] If it is determined that a target object exists in the image acquired by the fixed image acquisition, and the vibration signal and infrared signal are not effectively triggered, the feedback adjustment is performed below the trigger determination threshold.

[0023] Specifically, the steps for determining whether a target object exists in the acquired image include:

[0024] During device installation, an initial environmental image is acquired as a background image, an update cycle is set, and a background update process is triggered to update the background image.

[0025] After grayscale conversion and filtering for noise reduction of the acquired image, pixel-by-pixel difference is calculated between the acquired image and the background image. The absolute difference is then calculated, and a dynamic foreground region is selected by setting a dynamic threshold.

[0026] An adaptive threshold segmentation algorithm is used to calculate the target threshold for the dynamic foreground region and segment the foreground target region.

[0027] The brightness of the foreground target region is adjusted, and the size of the foreground target region is normalized. Then, the similarity with the target object database is calculated. The similarity between the foreground target region and the target object database consists of image similarity and feature similarity.

[0028] Specifically, the steps for determining whether a target object exists in the acquired image also include:

[0029] The presence of target objects in the acquired images is determined by calculating the pixel-level similarity between the foreground target region and images in the target object database at multiple scales.

[0030] The method for calculating the similarity of features to determine whether a target object exists in a captured image is as follows:

[0031] The shape and texture features of the foreground target region are extracted using a feature extraction algorithm. The shape feature similarity is obtained by calculating the Hu moment invariant Euclidean distance between the foreground target region and target images in the target database.

[0032] The texture histogram is extracted using a local binary mode. The texture similarity between the foreground target region and the target image in the target database is calculated using Bach distance. Finally, the feature similarity is obtained by weighted summation.

[0033] Configure image similarity threshold and feature similarity threshold. If the image similarity between a foreground target region and an image in the target object database is greater than the image similarity threshold, or the feature similarity is greater than the feature similarity threshold, then it is determined that there is a target object in the dynamic foreground region of the acquired image.

[0034] Specifically, the steps for automatically adjusting the brightness of the supplementary light based on the ambient light include:

[0035] During non-collection periods, the ambient light intensity is measured, a light threshold is configured, and the ambient light intensity is compared with the light threshold for determining whether a target object exists in the collected image. If the ambient light intensity is greater than the light threshold, the natural light is considered sufficient; otherwise, the natural light is considered insufficient.

[0036] For both trigger-based and stationary image capture, if sufficient natural light is determined, the image is captured directly using ambient light; if insufficient natural light is determined, the supplementary lighting adjustment process is initiated.

[0037] Set the base illumination brightness and illumination duration interval. Within the illumination duration interval of determining whether there is a target object in the acquired image, illuminate with the base illumination brightness and acquire the image. Obtain the grayscale mean and contrast of the acquired image.

[0038] Configure grayscale threshold and contrast threshold. If the average grayscale value of the acquired image is greater than the grayscale threshold and the contrast value is greater than the contrast threshold, it is determined to be a valid image; otherwise, it is determined to be an invalid image.

[0039] If the image is invalid, set the supplementary light brightness step size and gradually increase the supplementary light brightness step size on the base supplementary light brightness. Reacquire the image within the supplementary light duration interval, extract the dynamic region coordinates from the valid image and the invalid image through background difference, and replace the pixel values ​​of the dynamic region with the interpolation result of the pixel values ​​of the two images to merge and generate a valid image.

[0040] Specifically, the steps for automatically adjusting the brightness of the supplementary light according to the ambient light also include:

[0041] For video capture, if sufficient natural light is determined, ambient light is used directly for video capture; if insufficient natural light is determined, a dynamic supplemental lighting adjustment process is initiated.

[0042] Starting from the ambient light intensity, the brightness of the fill light is gradually and smoothly increased. The grayscale mean and contrast of the video frames are analyzed in real time. When the grayscale mean of the video frame is greater than the grayscale threshold and the contrast is greater than the contrast threshold, the dynamic fill light adjustment is stopped and the current fill light brightness is maintained to complete the video acquisition.

[0043] Specifically, the steps for obtaining the activity intensity of the target object and optimizing transmission include:

[0044] An adaptive contrast enhancement algorithm is enabled for the video frames received by the monitoring platform. A detection confidence threshold is configured to identify and label individual target objects in the video frames, and output the target object coordinates and timestamp information.

[0045] Configure a trajectory matching threshold to bind the ID of target objects in consecutive video frames and generate a motion trajectory sequence; configure a time window to calculate the average moving speed of the target objects within the time window based on the motion trajectory sequence.

[0046] The video frames are divided into grids, and the distribution of the number of targets within each grid is statistically analyzed. The difference in the number of targets between grids is mapped to a clustering index.

[0047] Configure the weight ratios of individual quantity, movement speed, and aggregation degree, calculate the activity intensity index by weighting, and smooth the activity intensity index within the configured smoothing window;

[0048] Configure low activity intensity threshold and high activity intensity threshold. When the smoothed activity intensity index is less than the low activity intensity threshold, trigger a pause video capture command.

[0049] Configure a single-use data quota. When the number of received video frames exceeds the single-use data quota and the activity intensity index exceeds the high activity intensity threshold, trigger the data compression mode.

[0050] The rate of change between video frames is calculated by pixel difference or feature vector difference, key frames are marked and redundant frames are removed, and the key frames are uploaded to the monitoring platform.

[0051] Specifically, the steps for adjusting the threshold for determining whether a target object exists in the acquired image include:

[0052] When the trigger judgment threshold is adjusted, the confidence level of the detected vibration signal and the confidence level of the infrared signal are recorded, and the false trigger count is increased.

[0053] Configure a false trigger threshold. If the false trigger count exceeds the false trigger threshold, it is determined to be a high-frequency false trigger, and the feedback adjustment is frozen.

[0054] Configure adjustment step size, including vibration adjustment step size and infrared adjustment step size. Increase the corresponding adjustment step size based on the current vibration threshold and infrared threshold respectively to adjust the judgment threshold upward.

[0055] When the feedback adjustment is triggered below the judgment threshold, the image that triggers the feedback adjustment below the judgment threshold is marked as a missed sample, and the missed detection count is increased.

[0056] Configure a missed detection threshold. When the missed detection count is greater than the preset missed detection threshold, it is determined that the detection sensitivity is insufficient. Based on the current vibration threshold and infrared threshold, the corresponding adjustment step size is reduced to lower the judgment threshold.

[0057] A target agent monitoring device based on AI visual recognition, comprising a trapping module, an intelligent power supply module, a visual recognition module, and an IoT transmission module;

[0058] The trapping module includes a biomimetic nest structure shell and a targeting agent, used to attract target objects into the trap and achieve mass extermination through the targeting agent;

[0059] The intelligent power supply module is used to provide power support through hybrid power supply and intelligent energy consumption regulation;

[0060] The visual recognition module is used to identify target objects and dynamically monitor their activity status;

[0061] The IoT transmission module is used for remote monitoring and information feedback of the target object.

[0062] The beneficial effects of this invention are:

[0063] This invention effectively solves the problem of false triggering caused by environmental noise interference in single sensors by extracting features from vibration and infrared signals, calculating confidence levels, and determining regional overlap, significantly improving the accuracy of image acquisition triggering. Based on a target object recognition algorithm using multi-scale image similarity and composite feature matching, combined with dynamic background updating and adaptive threshold segmentation, it overcomes the bottleneck of target recognition under low or complex lighting conditions, reducing the false detection rate. Through adaptive ambient light supplementation and pixel-level light intensity compensation and smooth brightness gradient adjustment, it ensures consistent image quality under different lighting conditions, avoiding feature loss due to insufficient brightness. The activity intensity-based transmission optimization strategy—trajectory tracking, gridded analysis, and keyframe screening—ensures the integrity of key termite infestation data while reducing redundant data transmission, lowering network load and system power consumption. Furthermore, the threshold feedback adjustment mechanism, through self-learning from false triggers and false detection samples, allows the detection sensitivity to dynamically adapt to the environment, reducing the false judgment rate caused by traditional fixed thresholds. Overall, this invention significantly improves the intelligence level and control efficiency of termite monitoring in complex environments, providing reliable technical support for early warning and precise extermination of termite infestations. Attached Figure Description

[0064] Figure 1 This is a flowchart of a target agent monitoring method based on AI visual recognition according to the present invention;

[0065] Figure 2 This is a flowchart illustrating the specific steps involved in triggering the camera to capture images according to the present invention.

[0066] Figure 3 This is a flowchart illustrating the specific steps involved in determining whether a target object exists in a captured image according to the present invention.

[0067] Figure 4 A flowchart illustrating the specific steps of this invention for obtaining the activity intensity of a target object and optimizing its transmission.

[0068] Figure 5 This is a schematic diagram of the structure of a target agent monitoring device based on AI visual recognition according to the present invention. Detailed Implementation

[0069] Example 1

[0070] Please see Figure 1This embodiment introduces a target-targeting agent monitoring method based on AI visual recognition. The target of this invention is insects, including but not limited to social insects such as termites and ants. When the target is termites, a biomimetic nest structure shell is adopted, with a slow-release pheromone bait stick inside. The bait stick attracts termites to feed by releasing termite pheromones that cause the target termite species to gather. The pheromone bait stick is loaded with a low-toxicity biodegradable bait, which is transmitted to the core of the nest through feeding behavior to achieve mass extermination. The attraction radius covers a range of 5 to 10 meters, combining targeted attraction and environmentally friendly extermination functions.

[0071] In this embodiment, a biomimetic nest structure shell is used. This shell is injection molded from a high-polymer composite material, and its surface simulates the rough texture and arched structure of a natural termite nest, possessing crush-resistant and waterproof properties. The inner wall of the shell is provided with spiral channels, the width of which matches the target termite species, guiding the termites along the channels into the core trapping area. At the same time, the structure's light-blocking properties reduce their alertness. The built-in slow-release pheromone trap targets target termite species such as subterranean termites and naked termites, using a combination of artificially synthesized aggregation pheromones such as farnesol derivatives and trace pheromones such as terpenoid compounds. The traps are controlled by gas chromatography and mass spectrometry. By optimizing the ratio using technology to ensure the attraction rate for specific ant species, pheromones are encapsulated in biodegradable microcapsules made of chitosan. The microcapsules gradually dissolve and release pheromones within the termite activity temperature range through a temperature-triggered release mechanism. The pheromone bait stick is loaded with environmentally friendly bait and supplemented with natural carriers such as starch and cellulose. After worker ants feed on it, they transfer the agent to larvae, queens, and kings through oral secretions. The agent is spread by utilizing the termite community feeding behavior. Through biomimetic structural design, precise pheromone ratio, and low-toxicity slow-release technology, targeted trapping and nest-level extermination of the target ant species can be achieved.

[0072] This device employs a hybrid power supply system combining solar panels and lithium batteries. It automatically charges when there is sufficient sunlight and automatically enters sleep mode when not in operation, achieving low power consumption and long-lasting battery life. In this embodiment, the solar panels are treated with an anti-fouling coating, allowing them to automatically charge the lithium batteries under sunlight. The lithium batteries are adaptable to a wide operating temperature range and integrate an intelligent adjustment chip that automatically optimizes charging efficiency based on light intensity, ensuring continuous operation even in the absence of sunlight. In the non-operational state, sensors detect the activity of target objects. When there is no activity, the device automatically enters sleep mode, retaining only basic monitoring functions, significantly reducing overall power consumption and achieving low power consumption and long-lasting battery life.

[0073] The system uses a camera to capture images at regular intervals, and AI algorithms analyze and identify the images to determine if a target object is present. It also integrates vibration and infrared sensors to trigger the camera to capture images, determining whether a target object is detected to reduce unnecessary power consumption and eliminate interference from non-target organisms. When a target object is detected, video capture is automatically triggered, recording the target object's feeding, movement, and other behaviors for analysis of group activity and population changes. During image and video capture, the system automatically adjusts the supplementary lighting brightness according to the ambient light to avoid interfering with the target object's behavior. In low light or at night, it automatically switches to infrared mode to ensure image clarity. The system analyzes images locally in real time, waking up the video capture and communication link only when a target object is detected. In non-working states, the camera and processor enter sleep mode for low-power operation.

[0074] Please see Figure 2 Preferably, the specific steps to trigger the camera to capture images include:

[0075] The vibration sensor monitors the vibration signal of the entire circumference of the device shell in real time, while the infrared sensor synchronously monitors the infrared signal of the heat source within the device's internal radius, forming a multi-dimensional detection signal.

[0076] The vibration and infrared signals are preprocessed to remove environmental noise by using a bandpass filter and normalization to highlight the regular vibration signals generated by the crawling and feeding of the target object.

[0077] For infrared signals, a Gaussian mixture model is used to distinguish between background heat sources and dynamic heat sources, extract the temperature gradient and movement speed of moving targets, and exclude jumping targets, such as beetles.

[0078] The confidence levels of the preprocessed vibration and infrared signals of the target object are calculated. The vibration signal confidence level is calculated based on a weighted fusion of effective vibration frequency and characteristics, while the infrared signal confidence level is calculated based on a proportional mapping between temperature gradient and moving speed. For example, the vibration signal confidence level... The calculation formula is:

[0079]

[0080] in, The number of vibration signal acquisitions. This represents the number of signals whose amplitude exceeds a preset amplitude threshold after bandpass filtering. The amplitude threshold is used to filter out environmental micro-vibration interference and retain the effective mechanical vibration signals generated by the target object's movement. The cosine similarity between the vibration signal frequency and the target object's characteristic frequency is used. The target object's characteristic frequency is the dominant vibration frequency of common borer species. This is used to quantify the degree of matching between the signal spectrum and the known activity frequency characteristics of the target object, eliminating interference from non-target frequencies. The upper limit of the preset characteristic frequency is used to limit the cutoff frequency of the bandpass filter and avoid high-frequency noise affecting the confidence calculation.

[0081] Infrared signal confidence The calculation formula is:

[0082]

[0083] in, The temperature difference between the heat source and the environment For the range of ambient temperature fluctuations, To measure the speed at which the heat source moves, This represents the typical crawling speed of the target object.

[0084] Configure judgment thresholds, including vibration threshold and infrared threshold. When the confidence level of the vibration signal is greater than the vibration threshold, the vibration signal is determined to be effectively triggered. When the confidence level of the infrared signal is greater than the infrared threshold, the infrared signal is determined to be effectively triggered.

[0085] When both vibration and infrared signals are effectively triggered, the internal detection area of ​​the device is divided according to the layout of the multimodal sensors. The vibration sensor determines the vibration source detection area by phase difference positioning, and the infrared sensor determines the heat source detection area by triangulation.

[0086] Determine whether the detection area effectively triggered by the vibration signal overlaps with the detection area effectively triggered by the infrared signal; if they overlap, trigger the camera to acquire an image to reduce invalid detections; otherwise, do not perform any processing.

[0087] When the camera is triggered to capture an image, the image inside the camera capture device is identified and determined to determine whether a target object exists in the captured image through edge computing.

[0088] Based on the identification results of whether a target object exists in the image captured by the camera, the image acquisition triggering conditions are adjusted accordingly, that is, the judgment threshold is adjusted accordingly, including:

[0089] If it is determined that there is no target object in the image acquired by the trigger, the trigger judgment threshold is adjusted accordingly, the confidence level of the vibration signal and the confidence level of the infrared signal detected this time are automatically recorded, and the false trigger count is increased, with an initial value of 0.

[0090] Configure a false trigger threshold. If the false trigger count exceeds the false trigger threshold, it is determined to be a high-frequency false trigger. Feedback adjustments are temporarily frozen to avoid frequent adjustments affecting stability in a short period of time.

[0091] Configure adjustment step size, including vibration adjustment step size and infrared adjustment step size. Increase the corresponding adjustment step size based on the current vibration threshold and infrared threshold respectively, and upload the adjusted judgment threshold and sensor parameters to the monitoring platform synchronously. Include the confidence formula parameters before and after adjustment, environmental data and historical false touch waveforms for manual review.

[0092] If it is determined that there is a target object in the image acquired by the fixed image, and the vibration signal and infrared signal are not effectively triggered, the feedback adjustment below the judgment threshold is triggered, the image is marked as a missed sample, and the missed count is increased, with an initial value of 0.

[0093] Configure a missed detection threshold. When the missed detection count exceeds the preset threshold, the detection sensitivity is deemed insufficient. The threshold is then lowered based on the current vibration threshold and infrared threshold, respectively, by reducing the corresponding adjustment step size. Relevant detection data, including missed detection images and recognition results, original sensor signal records, and a comparison of threshold parameters before and after adjustment, are sent to the monitoring platform for manual review.

[0094] Preferably, the specific steps for automatically adjusting the brightness of the supplementary light according to the ambient light include:

[0095] During non-collection periods, the system enters a low-power sleep mode to turn off the light-emitting components and measures the ambient light intensity using a photosensitive sensor.

[0096] Configure a light threshold. If the ambient light intensity is greater than the light threshold, such as light passing through soil cracks or diffuse reflection light on rainy days, then the natural light is considered sufficient; otherwise, the natural light is considered insufficient.

[0097] When triggering the camera to capture images or capturing stationary images, if sufficient natural light is determined, the image is captured directly using ambient light; if insufficient natural light is determined, a supplementary lighting adjustment process is initiated, i.e.:

[0098] Set the base illumination brightness, configure the illumination duration interval, and ensure that the illumination duration interval is shorter than the visual reaction time of the target object. Within the illumination duration interval, use the base illumination brightness to perform image acquisition and obtain the grayscale mean and contrast of the acquired image.

[0099] Configure image quality thresholds, including grayscale threshold and contrast threshold. If the average grayscale value of the acquired image is greater than the grayscale threshold and the contrast is greater than the contrast threshold, the acquired image is retained and determined to be a valid image to further determine whether the acquired image contains a target object; otherwise, it is determined to be an invalid image.

[0100] If an image is determined to be invalid, a supplementary lighting brightness step size is set, and the supplementary lighting brightness step size is gradually increased based on the base supplementary lighting brightness. Image acquisition is performed within the supplementary lighting interval. If an image is determined to be valid, the coordinates of the dynamic region in the invalid image are extracted through background subtraction, and the pixel values ​​of the dynamic region are replaced with the interpolated pixel values ​​of the valid and invalid images, and then merged to generate a valid image. For a single frame of invalid image, it is not directly deleted, but information is merged with the valid images of the preceding and following frames. Images with contrast less than or equal to the contrast threshold are removed during the merging process to reduce memory usage and retain the acquired information.

[0101] When capturing video, if sufficient natural light is determined, ambient light is used directly to capture the video; if insufficient natural light is determined, a dynamic supplemental lighting adjustment process is initiated, i.e.:

[0102] Starting with ambient light intensity, the supplementary lighting brightness is gradually and smoothly increased while simultaneously analyzing the average grayscale value and contrast of video frames in real time. When the average grayscale value of a video frame exceeds a grayscale threshold and the contrast ratio exceeds a contrast ratio threshold, dynamic supplementary lighting adjustment is stopped, and the current supplementary lighting brightness is maintained to complete video acquisition.

[0103] Please see Figure 3 Preferably, the specific steps for determining whether a target object exists in the acquired image include:

[0104] During device installation, initial environmental images are acquired as background images. Static features such as texture and structure of the device's inner wall are extracted to establish a background pixel distribution database. An update cycle is set, and the background update process is triggered during non-target object active periods. The background image is iteratively updated using an incremental Gaussian mixture model to eliminate long-term environmental interference such as dust accumulation and slight changes in lighting.

[0105] The effective image is sequentially converted to grayscale and then subjected to median filtering for noise reduction. The processed effective image is then compared pixel-by-pixel with the background image, and the absolute difference is calculated. A dynamic threshold is set, for example, the mean plus or minus two standard deviations. Pixels whose absolute difference falls outside the dynamic threshold range are retained; these pixels constitute the dynamic foreground region.

[0106] The dynamic foreground candidate region may contain residual background, such as reflections or translucent stains. An adaptive threshold segmentation algorithm is used to calculate the target threshold for the dynamic foreground region, thereby segmenting the foreground target region from the residual background and outputting a binary mask.

[0107] Different lighting variations can cause overexposure or underexposure of foreground objects, resulting in loss of detail. To address this, adaptive histogram equalization is used to adjust the brightness of the foreground object area while limiting the range of contrast enhancement to prevent feature distortion caused by overexposure or underexposure.

[0108] The segmented foreground target regions are normalized to ensure they match the image sizes in the database, eliminating size differences caused by distance, such as changes in image size as the target moves closer to or further from the lens, while preserving bounding box coordinates for subsequent localization. Then, similarity calculations are performed with a database of target images in memory, containing images of the same size but at different angles and shapes.

[0109] The similarity between the foreground target region and the target object database consists of image similarity and feature similarity.

[0110] Pixel similarity at a single scale is susceptible to local noise. Image similarity is determined by calculating the pixel-level similarity between the foreground target region and the target object image in the database at multiple scales, and selecting the highest value as the image similarity. This captures features from the overall outline to local details. Pixel-level similarity can use the multi-scale structural similarity index.

[0111] The feature similarity is calculated by using a feature extraction algorithm to extract the shape and texture features of the foreground target region. The shape feature similarity is obtained by calculating the Hu moment invariant Euclidean distance between the foreground target region and the target image in the target database. This method can stably describe key morphological features of the target object, such as beaded antennae and constricted abdomen, like the elliptical outline of Coleoptera insects.

[0112] Texture histograms are extracted using a local binary model. The texture similarity between the foreground target region and target images in the target database is calculated using Bach distance. Finally, a weighted summation is used to obtain the feature similarity. Subtle ripples on the target object's surface, such as segmental textures, are captured. Texture differences are quantified through histogram comparison, excluding non-target organisms with smooth surfaces, such as ants.

[0113] Configure image similarity thresholds and feature similarity thresholds to balance detection sensitivity and specificity. If the image similarity between a foreground target region and an image in the target database is greater than the image similarity threshold, or the feature similarity is greater than the feature similarity threshold, then the target is determined to exist in the dynamic foreground region of the valid image; otherwise, the target is determined not to exist in the dynamic foreground region of the valid image.

[0114] When video acquisition begins, real-time video frames are uploaded to the monitoring platform. The monitoring platform extracts target features from the received video frames, including the number of individuals, movement trajectory, and group aggregation degree, to obtain the target's activity intensity and optimize transmission: if the detected target activity intensity is lower than a preset threshold, a pause video acquisition command is automatically sent to save power; if the amount of transmitted data exceeds the daily quota, the data compression mode is triggered based on the target activity intensity, and key frames are transmitted first; during non-transmission periods, the IoT transmission module enters sleep mode, retaining only the wake-up signal listening function.

[0115] Please see Figure 4 Preferably, the specific steps for obtaining the activity intensity of the target object and optimizing transmission include:

[0116] The target monitoring environment may have low light or uneven lighting, which can cause the target outline to be blurred and affect the subsequent recognition accuracy. Therefore, the monitoring platform enables an adaptive contrast enhancement algorithm for the received video frames to improve the clarity of the target outline in low light environment by adjusting the brightness gain parameter.

[0117] To balance detection accuracy and efficiency under the limited computing power of edge devices and reduce false detections of non-target objects, a detection confidence threshold is configured, and a lightweight target detection model is used to identify individual target objects in video frames and annotate their bounding boxes, outputting target coordinates and timestamp information.

[0118] Analyzing video frames individually cannot capture the continuity of target behavior. To establish individual motion trajectories and analyze behavioral patterns, a multi-target tracking algorithm is employed, configuring trajectory matching thresholds, including spatial distance thresholds and time interval thresholds. Targets in consecutive video frames are ID-bound to generate motion trajectory sequences. The spatial distance threshold measures the spatial difference between target positions in different video frames. If the distance between a target's position in the current frame and its position in the previous frame falls within this threshold, it is considered the same target. The time interval threshold considers the continuity of targets in the temporal dimension, defining the time interval range for target appearances in consecutive video frames. For example, a time interval threshold of 0.5 seconds means that if a target appears in the current frame and its appearance in the previous frame is within 0.5 seconds of the current frame, they are considered the same target.

[0119] Single-frame speed is easily affected by noise, but the average speed over multiple consecutive frames can smoothly fluctuate and reflect the true movement state. A time window is configured to calculate the movement speed of the target object over multiple consecutive frames. Based on the movement trajectory sequence, the average movement speed of the target object is calculated within the time window length.

[0120] The degree of clustering of target objects is a key indicator for determining the hazard level, but direct counting makes it difficult to quantify spatial distribution characteristics. Therefore, video frames are divided into grids, and the distribution of the number of target objects within each grid is statistically analyzed.

[0121] The density variance algorithm is used to analyze the differences in the number of grids. Combined with the adjustable Gaussian kernel function parameters, the distribution differences are mapped to the clustering index in the interval [0,1]. The larger the value, the denser the clustering.

[0122] The activity intensity index is obtained by weighting the number of individuals, movement speed, and aggregation degree by configuring the weight ratios.

[0123] Real-time computing is susceptible to interference from random factors, resulting in frequent fluctuations. A smoothing window is configured to set the time range for filtering. Enabling the time window filtering function smooths the activity intensity index of real-time computing within the length of the smoothing window, reducing noise interference in a single frame.

[0124] Automatically identify periods with no target or weak activity to save energy; configure low activity intensity thresholds and high activity intensity thresholds; trigger a pause video capture command when the smoothed activity intensity index is less than the low activity intensity threshold.

[0125] To prioritize the transmission of high-value data under limited bandwidth, a single-data quota is configured. When the number of video frames received by the monitoring platform exceeds the single-data quota, and the smoothed activity intensity index exceeds the high activity intensity threshold, a data compression mode is triggered, prioritizing the transmission of key frames.

[0126] Configure a rate of change threshold. When acquiring video, calculate the rate of change between the current video frame and the previous video frame using pixel difference or feature vector difference. If the rate of change is greater than the threshold, mark the current video frame as a key frame; otherwise, it is determined to be a redundant frame and discarded. Set the interval between adjacent key frames, for example, at least 5 frames apart, to avoid transmitting excessively similar frames continuously. Upload the video frames marked as key frames to the monitoring platform to obtain the activity intensity of the target object.

[0127] Example 2

[0128] Please see Figure 5This embodiment introduces a target-targeting agent monitoring device based on AI visual recognition, comprising a trapping module, an intelligent power supply module, a visual recognition module, and an IoT transmission module. The trapping module is the core component for targeted trapping, and its core component includes a targeting agent, which is a key element in attracting target insects. In this embodiment, the targeting agent is a slow-release pheromone, which is placed on a trapping stick to form a slow-release pheromone trapping stick. The trapping module works closely with the targeting agent through a biomimetic nest-like shell structure. The shell simulates a natural habitat to reduce the target insects' alertness, while the trapping stick carrying the targeting agent enables precise release and efficient trapping of the pheromone, thereby attracting social insects such as termites and ants. The intelligent power supply module provides stable power support throughout the device, integrating low-power technology for long-lasting operation and ensuring continuous operation in complex environments. The visual recognition module relies on multi-sensor fusion perception and AI algorithm analysis to accurately identify targets and dynamically monitor their activity status, effectively eliminating interference from non-target organisms. The IoT transmission module handles data interaction between the device and a remote platform, enabling remote monitoring and information feedback of targets, achieving real-time uploading of monitoring data and precise issuance of control commands. These modules work together organically to achieve targeted monitoring and eradication of social insects, featuring strong targeting, environmental safety, and a high degree of intelligence.

[0129] The trapping module, as the core component for trapping and killing ants, includes a biomimetic nest structure shell and a targeting agent. The two work together to attract the target ant and then use the targeting agent to achieve mass extermination. The biomimetic nest structure shell is injection-molded from a high-polymer composite material. Its surface mimics the rough texture and arched structure of the target ant species' natural nest, possessing resistance to crushing and water immersion. The inner wall has spiral channels matching the body size of the target ant species, guiding the target into the core area while reducing its alertness through light-blocking properties. The slow-release pheromone trap, acting as the targeting agent, is embedded in the core area of ​​the biomimetic nest structure shell. This trap carries a compound pheromone for a specific target ant species; the slow-release pheromone, acting as the targeting agent, is a combination of artificially synthesized aggregation pheromones and trace pheromones, encapsulated in biodegradable chitosan microcapsules. Through a temperature-triggered mechanism, it is gradually released into an environmentally friendly bait within the target ant's activity temperature range. The environmentally friendly bait uses natural substances such as starch and cellulose as carriers, utilizing the target ant's feeding behavior to achieve toxic agent diffusion within the colony.

[0130] The intelligent power supply module is electrically connected to the trapping module, visual recognition module, and IoT transmission module. It provides power through hybrid power supply and intelligent energy consumption regulation, enabling low-power, long-lasting operation. It includes a solar panel, a lithium battery, and an intelligent regulation chip. The solar panel is treated with an anti-fouling coating, allowing it to charge the lithium battery under sunlight. The lithium battery is adaptable to a wide operating temperature range, and the intelligent regulation chip optimizes charging efficiency based on light intensity. It also integrates a sleep / wake-up mechanism, automatically entering sleep mode when there is no target activity signal, retaining only basic monitoring functions. It automatically charges when there is sufficient sunlight and reduces energy consumption when not in operation, ensuring continuous operation even in low-light environments.

[0131] The visual recognition module is electrically connected to the baiting module and the intelligent power supply module. Through image acquisition, sensor triggering, and AI algorithm analysis, it achieves accurate identification, behavior recording, and dynamic monitoring of target objects. It includes a camera, vibration sensor, infrared sensor, supplementary lighting component, and edge computing unit. The vibration sensor monitors the vibration signal of the device's outer casing in real time, while the infrared sensor simultaneously monitors the infrared signal of the internal heat source. The two sensors are integrated to trigger the camera to acquire images, reducing unnecessary power consumption. The camera can perform timed image acquisition and triggered video acquisition. During acquisition, the supplementary lighting component automatically adjusts its brightness according to the ambient light, switching to infrared mode in low light or at night. The edge computing unit has built-in AI algorithms that analyze images locally in real time, identifying target objects and recording their feeding, movement, and other behaviors to determine the activity level and population changes of the group.

[0132] The IoT transmission module is electrically connected to the vision recognition module and the intelligent power supply module, responsible for data transmission between the device and the external control system, enabling remote monitoring and information feedback. It receives target activity data transmitted by the vision recognition module, including images, videos, and quantity analysis results, and uploads it to the monitoring platform. Simultaneously, it can optimize the transmission strategy based on the target activity intensity and data volume, prioritizing the transmission of key frames and entering sleep mode during non-transmission periods. Furthermore, it can receive basic control parameters sent by the external control system and feed back the device's operating status information, such as power consumption, recognition results, and fault alerts, to the control system, enabling remote management and intelligent scheduling.

[0133] Working principle and its effects:

[0134] This invention uses a multimodal sensor to collect vibration and infrared signals in real time. After preprocessing such as bandpass filtering and Gaussian mixture model, features such as vibration frequency, temperature gradient and target movement speed are extracted. The signal confidence is calculated by weighted fusion and proportional mapping. The detection area is divided by phase difference positioning and triangulation. Image acquisition is triggered only when the vibration and infrared signals are effectively triggered and the detection area overlaps. This effectively filters environmental noise interference and significantly reduces the false trigger rate.

[0135] The algorithm extracts dynamic foreground from the acquired images through background subtraction, and identifies the target object by using a composite comparison algorithm that combines multi-scale pixel similarity with Hu rectangular features and local binary pattern texture features. At the same time, it automatically updates the background image during inactive periods, which solves the problem of target blurring under low light or complex lighting conditions and greatly improves the recognition accuracy.

[0136] In the supplementary lighting adjustment stage, the supplementary lighting brightness is dynamically adjusted according to the ambient light intensity. During image acquisition, pixel-level light intensity compensation is performed by evaluating the grayscale mean and contrast. During video acquisition, a smooth brightness gradient adjustment is used to ensure consistent image quality under different lighting conditions and avoid feature loss due to sudden brightness changes.

[0137] The monitoring platform tracks video frames to generate motion trajectories and calculates an activity intensity index using a gridded distribution statistics and time window smoothing algorithm. When the data volume exceeds the quota, key frames are selected for transmission based on the inter-frame change rate, significantly reducing redundant data transmission and lowering network load and power consumption while ensuring effective information. A threshold feedback adjustment mechanism based on self-learning of missed and false alarm samples dynamically optimizes vibration and infrared signal judgment thresholds, enabling detection sensitivity to adapt to environmental changes and effectively reducing false alarm and missed detection rates.

[0138] Overall, this invention constructs a target monitoring system with low false positive rate, high robustness, and strong adaptability, providing an efficient technical approach for early warning and precise extermination of ant infestations.

[0139] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for monitoring target agents based on AI visual recognition, characterized in that, include: Step S1: Real-time acquisition and fusion of device vibration signal and infrared signal, extraction of vibration frequency features, temperature gradient features and moving speed features, by configuring a judgment threshold to determine whether the vibration signal and infrared signal are both effectively triggered. If so, the detection area overlap analysis is performed based on the phase difference positioning of the vibration source and the triangulation measurement coordinates of the heat source. If there is an overlapping detection area, the camera is triggered to acquire images. Step S2: Perform background subtraction processing on the acquired image to extract the dynamic foreground region, segment the foreground target region, and determine whether there is a target object in the acquired image by combining multi-scale image similarity matching with composite similarity calculation of shape features and texture features. If there is a target object, video acquisition is performed; and the determination threshold is adjusted based on the target object recognition result. Step S3: After starting video acquisition, the video frames are uploaded to the monitoring platform. The monitoring platform is used to extract target features from the video frames, generate speed sequence data based on the target's movement trajectory tracking, and obtain the target's activity intensity by combining gridded distribution statistics and time window smoothing, and optimize the transmission. Step S4: During image and video acquisition, the brightness of the fill light is automatically adjusted according to the ambient light. During image acquisition, pixel-level light intensity compensation is performed based on grayscale mean and contrast evaluation. During video acquisition, a smooth brightness gradient adjustment is used until the preset image quality threshold is reached. The specific steps for triggering the camera to capture images include: The vibration sensor monitors the vibration signal of the device in real time, and the infrared sensor synchronously monitors the infrared signal of the heat source inside the device, forming a multi-dimensional detection signal. Preprocessing of vibration and infrared signals for target objects: The vibration signal is filtered by a bandpass filter to remove environmental noise and then normalized; the infrared signal is distinguished from the background and dynamic heat source based on a Gaussian mixture model, and the temperature gradient and moving speed of the moving target are extracted. The confidence level of the vibration signal is calculated by weighted fusion of effective vibration number and frequency characteristics, and the confidence level of the infrared signal is calculated by mapping the ratio of temperature gradient to moving speed. Configure judgment thresholds, including vibration threshold and infrared threshold. When the confidence level of the vibration signal is greater than the vibration threshold, the vibration signal is determined to be effectively triggered. When the confidence level of the infrared signal is greater than the infrared threshold, the infrared signal is determined to be effectively triggered. When both vibration and infrared signals are effectively triggered, the detection area is divided according to the layout of the multimodal sensors. The vibration sensor determines the vibration source detection area by phase difference positioning, and the infrared sensor determines the heat source detection area by triangulation. Determine whether the detection area effectively triggered by the vibration signal overlaps with the detection area effectively triggered by the infrared signal; if they overlap, trigger the camera to capture an image; otherwise, do not perform any processing. Based on the identification results of whether a target object exists in the image captured by the camera, the judgment threshold is adjusted accordingly, including: If it is determined that there is no target object in the image acquired by the trigger, the trigger determination threshold is adjusted accordingly. If it is determined that a target object exists in the image acquired by the fixed image acquisition, and the vibration signal and infrared signal are not effectively triggered, the feedback adjustment is performed below the trigger determination threshold.

2. The target agent monitoring method based on AI visual recognition as described in claim 1, characterized in that, The specific steps for determining whether a target object exists in the acquired image include: During device installation, an initial environmental image is acquired as a background image, an update cycle is set, and a background update process is triggered to update the background image. After grayscale conversion and filtering for noise reduction of the acquired image, pixel-by-pixel difference is calculated between the acquired image and the background image. The absolute difference is then calculated, and a dynamic foreground region is selected by setting a dynamic threshold. An adaptive threshold segmentation algorithm is used to calculate the target threshold for the dynamic foreground region and segment the foreground target region. The brightness of the foreground target region is adjusted, and the size of the foreground target region is normalized. Then, the similarity with the target object database is calculated. The similarity between the foreground target region and the target object database consists of image similarity and feature similarity.

3. The target agent monitoring method based on AI visual recognition as described in claim 2, characterized in that, The specific steps for determining whether a target object exists in the acquired image also include: The image similarity is determined by calculating the pixel-level similarity between the foreground target region and the images in the target object database at multiple scales; The feature similarity is calculated as follows: The shape and texture features of the foreground target region are extracted using a feature extraction algorithm; the similarity of the shape features is obtained by calculating the Hu moment invariant Euclidean distance between the foreground target region and the target image in the target database. The texture histogram is extracted using local binary mode. The texture similarity between the foreground target region and the target image in the target database is calculated using Bach distance. Finally, the feature similarity is obtained by weighted summation. Configure image similarity threshold and feature similarity threshold. If the image similarity between a foreground target region and an image in the target object database is greater than the image similarity threshold, or the feature similarity is greater than the feature similarity threshold, then it is determined that there is a target object in the dynamic foreground region of the acquired image.

4. The target agent monitoring method based on AI visual recognition as described in claim 1, characterized in that, The specific steps for automatically adjusting the supplementary light brightness according to ambient light include: During non-collection periods, the ambient light intensity is measured, a light threshold is configured, and the ambient light intensity is compared with the light threshold. If the ambient light intensity is greater than the light threshold, the natural light is determined to be sufficient; otherwise, the natural light is determined to be insufficient. For both trigger-based and stationary image capture, if sufficient natural light is determined, the image is captured directly using ambient light; if insufficient natural light is determined, the supplementary lighting adjustment process is initiated. Set the base fill light brightness and fill light duration interval, and within the fill light duration interval, fill light with the base fill light brightness and acquire images, and obtain the grayscale mean and contrast of the acquired images. Configure grayscale threshold and contrast threshold. If the average grayscale value of the acquired image is greater than the grayscale threshold and the contrast value is greater than the contrast threshold, it is determined to be a valid image; otherwise, it is determined to be an invalid image. If the image is invalid, set the supplementary light brightness step size and gradually increase the supplementary light brightness step size on the base supplementary light brightness. Reacquire the image within the supplementary light duration interval, extract the dynamic region coordinates from the valid image and the invalid image through background difference, and replace the pixel values ​​of the dynamic region with the interpolation result of the pixel values ​​of the two images to merge and generate a valid image.

5. The target agent monitoring method based on AI visual recognition as described in claim 4, characterized in that, The specific steps for automatically adjusting the supplementary light brightness according to ambient light also include: For video capture, if sufficient natural light is determined, ambient light is used directly for video capture; if insufficient natural light is determined, a dynamic supplemental lighting adjustment process is initiated. Starting from the ambient light intensity, the brightness of the fill light is gradually and smoothly increased. The grayscale mean and contrast of the video frames are analyzed in real time. When the grayscale mean of the video frame is greater than the grayscale threshold and the contrast is greater than the contrast threshold, the dynamic fill light adjustment is stopped and the current fill light brightness is maintained to complete the video acquisition.

6. The target agent monitoring method based on AI visual recognition as described in claim 1, characterized in that, The specific steps for obtaining the target object's activity intensity and optimizing transmission include: An adaptive contrast enhancement algorithm is enabled for the video frames received by the monitoring platform. A detection confidence threshold is configured to identify and label individual target objects in the video frames, and output the target object coordinates and timestamp information. Configure a trajectory matching threshold to bind the ID of target objects in consecutive video frames and generate a motion trajectory sequence; configure a time window to calculate the average moving speed of the target objects within the time window based on the motion trajectory sequence. The video frames are divided into grids, and the distribution of the number of targets within each grid is statistically analyzed. The difference in the number of targets between grids is mapped to a clustering index. Configure the weight ratios of individual quantity, movement speed, and aggregation degree, calculate the activity intensity index by weighting, and smooth the activity intensity index within the configured smoothing window; Configure low activity intensity threshold and high activity intensity threshold. When the smoothed activity intensity index is less than the low activity intensity threshold, trigger a pause video capture command. Configure a single-use data quota. When the number of received video frames exceeds the single-use data quota and the activity intensity index exceeds the high activity intensity threshold, trigger the data compression mode. The rate of change between video frames is calculated by pixel difference or feature vector difference, key frames are marked and redundant frames are removed, and the key frames are uploaded to the monitoring platform.

7. The target agent monitoring method based on AI visual recognition as described in claim 1, characterized in that, The specific steps for adjusting the judgment threshold include: When the trigger judgment threshold is adjusted, the confidence level of the detected vibration signal and the confidence level of the infrared signal are recorded, and the false trigger count is increased. Configure a false trigger threshold. If the false trigger count exceeds the false trigger threshold, it is determined to be a high-frequency false trigger, and the feedback adjustment is frozen. Configure adjustment step size, including vibration adjustment step size and infrared adjustment step size. Increase the corresponding adjustment step size based on the current vibration threshold and infrared threshold respectively to adjust the judgment threshold upward. When the feedback adjustment is triggered below the judgment threshold, the image that triggers the feedback adjustment below the judgment threshold is marked as a missed sample, and the missed detection count is increased. Configure a missed detection threshold. When the missed detection count is greater than the preset missed detection threshold, it is determined that the detection sensitivity is insufficient. Based on the current vibration threshold and infrared threshold, the corresponding adjustment step size is reduced to lower the judgment threshold.

8. A target agent monitoring device based on AI visual recognition, implemented based on any one of claims 1-7, characterized in that, It consists of a trapping module, an intelligent power supply module, a visual recognition module, and an IoT transmission module; The trapping module includes a biomimetic nest structure shell and a targeting agent, used to attract target objects into the trap and achieve mass extermination through the targeting agent; The intelligent power supply module is used to provide power support through hybrid power supply and intelligent energy consumption regulation; The visual recognition module is used to identify target objects and dynamically monitor their activity status; The IoT transmission module is used for remote monitoring and information feedback of the target object.

Images