Wildlife channel monitoring method based on infrared sensing and AI recognition

Abstract

This invention relates to the field of intelligent wildlife monitoring technology, and provides a wildlife channel monitoring method based on infrared sensing and AI recognition. The method includes triggering high-resolution infrared image acquisition when an infrared sensing module detects heat changes within a channel exceeding a threshold; identifying animal targets in an image sequence using an AI target detection model, extracting motion areas, and generating horizontal and vertical motion vector maps; calculating motion intensity distribution based on the vector maps, and generating a dynamic weight map based on a database of species activity patterns; and spatiotemporally fusing the dynamic weight map with infrared images of the current frame and historical frames to output monitoring results with species classification and activity trajectories, which are then uploaded to a data center. This invention enables precise monitoring, intelligent analysis, and real-time early warning of wildlife activities, improving the automation level and data value of wildlife channel monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent wildlife monitoring technology, and more specifically, to a wildlife channel monitoring method based on infrared sensing and AI recognition. Background Art

[0002] With growing awareness of wildlife conservation, wildlife corridor monitoring technology has become a key research area in ecological conservation. Traditional wildlife monitoring methods primarily rely on manual field observation and simple camera recording, which present numerous limitations. Manual observation is not only labor-intensive and time-consuming, but also difficult to achieve continuous monitoring over long periods of time and is easily affected by natural conditions such as weather and terrain. While traditional camera recording can capture animal activity to a certain extent, it is ineffective at night or in low-light conditions and cannot accurately identify animal species and behavioral patterns. While infrared sensing and image recognition technologies have seen some application in wildlife monitoring in recent years, these technologies are mostly limited to simple triggering and preliminary identification, incapable of in-depth analysis of animal behavior and real-time monitoring. Furthermore, existing technologies suffer from shortcomings in data transmission and processing, resulting in low data transmission efficiency, a lack of timely feedback on monitoring results, and a lack of effective adaptive update mechanisms for unidentified species.

[0003] In the process of implementing the embodiments of the present invention, it was found that there are at least the following problems or defects in the existing technology: First, the existing monitoring technology has poor monitoring effect at night or in low light conditions, and cannot accurately capture the activities of animals; second, the existing image recognition technology does not analyze animal behavior deeply enough and cannot provide detailed movement trajectories and behavior patterns; third, the existing technology has problems of low efficiency and unstable transmission quality during the data transmission process, and cannot meet the needs of real-time monitoring; finally, the existing technology lacks an effective adaptive update mechanism when facing unidentified species, and cannot update the model in time to adapt to new monitoring needs. Summary of the Invention

[0004] The present invention provides a wildlife channel monitoring method based on infrared sensing and AI recognition, comprising:

[0005] When the infrared sensing module detects that the heat change in the channel exceeds a first preset threshold, it triggers high-resolution infrared image acquisition to generate an infrared image sequence of the wildlife channel;

[0006] Performing animal target recognition on each frame of the infrared image sequence using an AI target detection model, extracting the motion area of ​​the animal target in each frame and generating a horizontal motion vector diagram and a vertical motion vector diagram relative to the previous frame;

[0007] Calculating the motion intensity distribution of the animal target in each frame of the image based on the horizontal motion vector diagram and the vertical motion vector diagram, and generating a dynamic weight map in combination with the species activity pattern database;

[0008] The dynamic weight map is spatially and temporally fused with the infrared images of the current frame and the historical frames, and the monitoring results with species classification and activity trajectory are output and uploaded to the data center via a wireless network.

[0009] Furthermore, triggering high-resolution infrared image acquisition includes:

[0010] Monitor pixel-level thermal changes in the infrared sensor module in real time. When the thermal change in the same area for a preset number of consecutive frames exceeds the ambient noise threshold, the infrared camera is activated to shoot at the highest resolution and switches to low-power mode when there is no activity.

[0011] Furthermore, the construction of the AI ​​target detection model includes:

[0012] A lightweight convolutional neural network is used as the backbone network, and a thermal radiation intensity normalization layer is added to the infrared image input layer to linearly map the original thermal radiation value to a standard grayscale range.

[0013] When loading a pre-trained species classification model through transfer learning, freeze the parameters of the feature extraction layer except the classification layer;

[0014] When deploying edge computing devices, a model quantization engine is used to quantize the network at a low bit width to achieve real-time detection of a preset frame rate. The detection threshold is set to a preset confidence threshold and a preset intersection-to-union ratio threshold.

[0015] Furthermore, generating the horizontal motion vector map and the vertical motion vector map includes:

[0016] Perform pixel-level matching on the motion areas of the same animal target in two adjacent frames of images, and calculate the horizontal and vertical displacement vectors based on the matching results;

[0017] The displacement vector is mapped to the entire image according to the spatial position to generate a horizontal motion vector map and a vertical motion vector map respectively.

[0018] Furthermore, the calculation of the motion intensity distribution includes:

[0019] According to the displacement of each pixel in the horizontal motion vector map and the movement speed threshold of the corresponding species, high activity area and low activity area are divided;

[0020] The same processing is performed on the longitudinal motion vector map, and the transverse and longitudinal activity areas are superimposed to generate a two-dimensional motion intensity distribution map.

[0021] Furthermore, the generation of the dynamic weight map includes:

[0022] Performing similarity matching on the motion intensity distribution map and typical activity templates in a species activity pattern database, and adjusting the weight coefficient of each region according to the matching result;

[0023] The weight coefficients are smoothed in time series to generate a dynamic weight map for spatiotemporal fusion.

[0024] Furthermore, the space-time fusion includes:

[0025] The animal target area of ​​the current frame is locally enhanced according to the dynamic weight map, and the background area is denoised using the multi-frame mean of historical frames;

[0026] The species classification labels are retained during the fusion process, and the motion trajectories of the same target in consecutive frames are connected into a complete path.

[0027] Furthermore, the method further includes a model adaptive updating step:

[0028] When an unidentified species is detected for a preset number of consecutive frames and the target area ratio exceeds a preset area threshold, the image acquisition device is automatically triggered to save the target area and adjacent frame images;

[0029] After performing background subtraction and histogram equalization preprocessing on the collected images, pseudo labels are generated through semi-supervised learning, and the species classification model is updated using an online knowledge distillation algorithm;

[0030] During incremental training, the feature extraction layer is frozen and only the classification layer parameters are updated. After training is completed, a model validation test is performed, and an updated version is deployed when the accuracy improvement exceeds the preset accuracy threshold.

[0031] Furthermore, the data compression and transmission step includes:

[0032] Inter-frame compression coding is used for static background areas, and the compression rate is dynamically adjusted according to the wireless network signal strength;

[0033] When animal motion trajectory data is converted into vector format, each trajectory point contains timestamp, thermal radiation intensity, and motion direction data, and differential coding is used to compress the direction angle;

[0034] Prioritize uploading classification information containing preset protected species, trigger emergency transmission mode when abnormal movement patterns are detected, and improve the keyframe compression quality to above the preset threshold.

[0035] Furthermore, abnormal behavior analysis includes:

[0036] Calculate the steering angles formed by adjacent trajectory points in the motion trajectory. If the cumulative steering angle exceeds an angle threshold within a preset time period and the number of occurrences of a single steering angle exceeding a sub-angle threshold reaches a preset number, it is determined to be wandering behavior;

[0037] Speed ​​mutation detection uses sliding window variance analysis to generate a poaching warning when the speed change exceeds the species average speed threshold within a preset time window and the frequency of occurrence reaches a preset frequency;

[0038] When analyzing multi-species trajectories, the degree of overlap of activity areas of different species is detected through spatial topology maps. When the proportion of overlapping area exceeds the area threshold and the duration exceeds the time threshold, the habitat conflict level is comprehensively determined in combination with multimodal sensor data.

[0039] The above-mentioned embodiments according to the present invention have at least the following beneficial effects: the present invention can achieve accurate monitoring and efficient analysis of wildlife passages. Through the combination of the infrared sensing module and high-resolution infrared image acquisition, it is possible to clearly capture the activity images of wild animals at night or in low light conditions, ensuring the continuity and integrity of monitoring. The introduction of the AI ​​target detection model makes the identification of animal targets more accurate, can quickly distinguish different species and extract their movement areas, generate detailed motion vector maps, and then calculate the motion intensity distribution, providing rich data for studying animal behavior. The generation of dynamic weight maps and the application of spatiotemporal fusion technology can further enhance the accuracy and reliability of monitoring results, can clearly display the activity trajectories and behavior patterns of animals, and provide strong support for wildlife protection and ecological research.

[0040] Furthermore, the present invention can also achieve adaptive model updates and efficient data transmission. When an unidentified species is detected, the system can automatically trigger image acquisition and preprocessing, and update the species classification model through semi-supervised learning and online knowledge distillation algorithms to ensure the accuracy and adaptability of the model. Data compression and transmission technology can dynamically adjust the compression rate based on the wireless network signal strength, prioritize uploading critical data, and trigger emergency transmission mode when abnormal motion patterns are detected. This improves the efficiency and reliability of data transmission, reduces data transmission costs, and reduces network bandwidth usage, ensuring that monitoring data can be uploaded to the data center in a timely and accurate manner, providing strong support for subsequent analysis and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0042] Figure 1A flowchart of a wildlife channel monitoring method based on infrared sensing and AI recognition provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0043] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0044] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0045] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0046] Reference below Figure 1 , Figure 1 This is a flow chart of a wildlife channel monitoring method based on infrared sensing and AI recognition provided by one embodiment of the present invention. Figure 1 As shown, a wildlife channel monitoring method based on infrared sensing and AI recognition includes:

[0047] S1. When the infrared sensing module detects that the heat change in the channel exceeds a first preset threshold, it triggers high-resolution infrared image acquisition to generate an infrared image sequence of the wildlife channel;

[0048] S2. Performing animal target recognition on each frame of the infrared image sequence using an AI target detection model, extracting the motion area of ​​the animal target in each frame and generating a horizontal motion vector diagram and a vertical motion vector diagram relative to the previous frame;

[0049] S3, calculating the motion intensity distribution of the animal target in each frame image based on the horizontal motion vector diagram and the vertical motion vector diagram, and generating a dynamic weight map in combination with a species activity pattern database;

[0050] S4. Perform spatiotemporal fusion of the dynamic weight map with the infrared images of the current frame and the historical frames, output monitoring results with species classification and activity trajectory, and upload them to the data center via a wireless network.

[0051] It should be noted that when the infrared sensing module detects a heat change within the channel exceeding a first preset threshold, it triggers high-resolution infrared image acquisition, generating a sequence of infrared images of the wildlife channel. The infrared sensing module here is a device capable of sensing heat changes. It determines whether wildlife is passing through by detecting heat changes within the channel. A heat change exceeding the first preset threshold indicates that an animal has entered the monitoring area, triggering high-resolution infrared image acquisition. High-resolution infrared image acquisition involves using an infrared camera to capture images at high resolution, generating a sequence of infrared images containing wildlife information. These image sequences contain information about the animal's thermal radiation, clearly showing the animal's outline and position, providing basic data for subsequent analysis.

[0052] Specifically, the pixel-level heat change of the infrared sensing module refers to the smallest unit of heat change that the module can detect. This change can be detected by the module's sensor as heat changes at different locations within the channel. The first preset threshold is a set heat change value. When the detected heat change exceeds this value, the system determines that an animal has entered the monitoring area. High-resolution infrared image acquisition uses an infrared camera to capture images at high resolution, generating an infrared image sequence containing wildlife information. High resolution here refers to the level of image detail the camera can capture, typically measured in pixels. For example, a high-resolution infrared camera can capture images at 1920×1080 pixels, which means it can provide clear images of animals, facilitating subsequent analysis and identification. An infrared image sequence is a series of continuously captured infrared images that record the activities of animals within the channel, providing rich data for subsequent analysis.

[0053] Preferably, the infrared sensing module can use a thermopile sensor, which can convert heat changes into electrical signals, thereby realizing accurate detection of heat changes in the channel. The first preset threshold value can be adjusted according to the size of the animal and the ambient temperature of the channel. For example, a lower threshold value can be set for small animals, and a higher threshold value can be set for large animals. High-resolution infrared image acquisition can be achieved by setting the resolution parameters of the infrared camera. For example, the resolution of the camera can be set to 1920×1080 pixels to ensure that clear images of animals can be captured. When generating an infrared image sequence, the shooting frequency of the camera can be set, for example, 30 frames per second, to ensure that the continuous activity process of the animal can be recorded.

[0054] In some embodiments, triggering high-resolution infrared image acquisition includes:

[0055] Monitor pixel-level thermal changes in the infrared sensor module in real time. When the thermal change in the same area for a preset number of consecutive frames exceeds the ambient noise threshold, the infrared camera is activated to shoot at the highest resolution and switches to low-power mode when there is no activity.

[0056] It should be noted that the infrared sensing module monitors pixel-level heat changes in real time. When the heat change amplitude in the same area in a preset number of consecutive frames exceeds the environmental noise threshold, the infrared camera is activated to shoot at the highest resolution and switches to low-power mode when there is no activity. The pixel-level heat change here refers to the smallest unit of heat change that the infrared sensing module can detect, usually measured in pixels. The environmental noise threshold is a baseline value set to distinguish between heat changes generated by real animal activities and environmental background noise, such as wind and shaking vegetation. The preset number of frames refers to the number of frames that need to meet the heat change threshold during continuous monitoring to ensure that continuous animal activity is detected rather than accidental interference. Low-power mode is an energy-saving state. When no animal activity is detected, the infrared camera will switch to this mode to reduce energy consumption and extend the service life of the device.

[0057] Specifically, the infrared sensing module detects pixel-level thermal changes through the sensor's sensitivity, enabling it to detect thermal changes at each pixel. The ambient noise threshold is typically set based on actual background thermal fluctuations. For example, in windy environments, the ambient noise threshold can be set relatively high to avoid false triggering. The preset frame rate can be determined based on the animal's average movement speed and the camera's frame rate. For example, if the animal's average movement speed is fast, the preset frame rate can be set lower, while if it is fast, the preset frame rate can be set higher. Low-power mode is a power-saving state that the camera enters when no activity is detected. This significantly reduces power consumption, typically achieved by lowering the camera's sampling rate or disabling certain features. For example, if the thermal fluctuations in the same area across 10 consecutive frames exceed the ambient noise threshold, animal activity is detected, activating the infrared camera to capture at its highest resolution. When there is no activity, the camera switches to low-power mode to conserve energy.

[0058] Preferably, the infrared sensing module can use a highly sensitive thermopile sensor, which can generate a measurable electrical signal under tiny heat changes. The environmental noise threshold can be determined by sampling the background heat fluctuations of the monitored environment for a period of time. For example, a suitable threshold can be set by analyzing the background heat changes in different time periods within a month. The preset number of frames can be set according to the average movement speed of the animal and the frame rate of the camera. For example, for small, fast-moving animals, the preset number of frames can be set to 5 frames, while for large, slow-moving animals, it can be set to 10 frames. Low power consumption mode can be achieved by setting the power management parameters of the camera. For example, when no heat change is detected for 30 consecutive seconds, the camera automatically switches to low power consumption mode. At this time, the power consumption of the camera can be reduced to less than 10% of the normal power consumption, thereby effectively extending the battery life of the device.

[0059] In some embodiments, the construction of the AI ​​target detection model includes:

[0060] A lightweight convolutional neural network is used as the backbone network, and a thermal radiation intensity normalization layer is added to the infrared image input layer to linearly map the original thermal radiation value to a standard grayscale range.

[0061] When loading a pre-trained species classification model through transfer learning, freeze the parameters of the feature extraction layer except the classification layer;

[0062] When deploying edge computing devices, a model quantization engine is used to quantize the network at a low bit width to achieve real-time detection of a preset frame rate. The detection threshold is set to a preset confidence threshold and a preset intersection-to-union ratio threshold.

[0063] It should be noted that the construction of the AI ​​object detection model includes a lightweight convolutional neural network as the backbone network. A thermal radiation intensity normalization layer is added to the infrared image input layer to linearly map the raw thermal radiation values ​​to a standard grayscale range. When loading a pre-trained species classification model through transfer learning, the parameters of the feature extraction layer, except for the classification layer, are frozen. When deployed on edge computing devices, a model quantization engine is used to quantize the network to a low bit width, enabling real-time detection at a preset frame rate. The detection threshold is set to a preset confidence threshold and an intersection-over-union threshold. The lightweight convolutional neural network here refers to an optimized neural network structure designed to reduce computational effort and model size while maintaining high accuracy. The thermal radiation intensity normalization layer is a network layer used to normalize the thermal radiation values ​​in the infrared image to a fixed grayscale range for subsequent processing. Transfer learning is a machine learning method that accelerates training and improves model performance by transferring the parameters of a pre-trained model to a new task. The model quantization engine is a tool used to convert model parameters from floating-point numbers to a low-bit-width quantized representation to reduce model storage requirements and computational complexity.

[0064] Specifically, lightweight convolutional neural networks improve the efficiency of the model by reducing the number of parameters and computational complexity in the network. For example, lightweight architectures such as MobileNet or ShuffleNet can be used. The function of the thermal radiation intensity normalization layer is to linearly map the thermal radiation values ​​in the infrared image from the original range, such as 0 to 1000, to a standard grayscale range, such as 0 to 255. This ensures that image data from different devices or different environments are comparable. In transfer learning, the pre-trained species classification model is a model trained on a large-scale dataset. By freezing the feature extraction layer parameters and only updating the classification layer parameters, it can quickly adapt to new species classification tasks. The model quantization engine can quantize the model parameters from 32-bit floating point numbers to 8-bit integers, thereby significantly reducing the model size and computational complexity. The preset frame rate refers to the number of image frames that the model can process on the edge computing device, such as 30 frames per second. The detection threshold includes a preset confidence threshold and a preset intersection-over-union threshold. The preset confidence threshold is used to determine the model's confidence in the detection results and is usually set between 0.5 and 0.9; the preset intersection-over-union threshold is used to evaluate the accuracy of the detection box and is usually set to 0.5.

[0065] Preferably, constructing a lightweight convolutional neural network can include selecting an appropriate network architecture, such as MobileNetV2, which reduces computational complexity through depthwise separable convolutions. The specific implementation of the thermal radiation intensity normalization layer can be to subtract the minimum thermal radiation value from each pixel, divide it by the difference between the maximum and minimum values, and then multiply by 255. In transfer learning, the pre-trained model can be a ResNet model trained on the ImageNet dataset. By freezing the parameters of the first few layers and fine-tuning only the final classification layer, it can be quickly adapted to new species classification tasks. The model quantization engine can use tools such as TensorFlow Lite or ONNX Runtime to quantize model parameters from 32-bit floating point to 8-bit integers while maintaining model accuracy. When deployed on edge computing devices, model performance can be optimized by adjusting the model input size and quantization parameters, for example, resizing the input image to 224×224 pixels and setting the quantization parameters to 8-bit integers. The preset frame rate can be set based on the computing power of the device. For example, it can be set to 60 frames per second for high-performance edge computing devices and 15 frames per second for low-power devices. The detection threshold can be adjusted according to the specific application scenario. For example, in a scenario with high detection accuracy requirements, the preset confidence threshold can be set to 0.8 and the preset intersection-over-union threshold can be set to 0.6.

[0066] In some embodiments, generating the horizontal motion vector map and the vertical motion vector map includes:

[0067] Perform pixel-level matching on the motion areas of the same animal target in two adjacent frames of images, and calculate the horizontal and vertical displacement vectors based on the matching results;

[0068] The displacement vector is mapped to the entire image according to the spatial position to generate a horizontal motion vector map and a vertical motion vector map respectively.

[0069] It should be noted that the process of generating horizontal and vertical motion vectors is to perform pixel-level matching of the moving areas of the same animal target in two adjacent frames. Based on the matching results, the horizontal and vertical displacement vectors are calculated. These displacement vectors are then mapped to the entire image according to their spatial position to generate the horizontal and vertical motion vectors, respectively. Pixel-level matching here refers to precise matching at the pixel level of the image to determine the positional changes of the animal target between two adjacent frames. The horizontal and vertical displacement vectors represent the horizontal and vertical displacements of the animal target, respectively. These vectors can reflect the direction and speed of the animal's movement.

[0070] Specifically, pixel-level matching is achieved by comparing the grayscale value or features of each pixel in two adjacent frames. For example, the optical flow method can be used to calculate the displacement between pixels. The horizontal displacement vector and the vertical displacement vector are obtained by calculating the displacement of each pixel in the horizontal and vertical directions. These displacement vectors can be represented as arrows pointing from a pixel point to its corresponding point in the next frame of the image. Spatial position mapping refers to placing these displacement vectors at the corresponding positions of the image to form a complete vector field. The horizontal motion vector diagram and the vertical motion vector diagram respectively represent the movement of the animal target in the horizontal and vertical directions. These diagrams can be used to further analyze the movement patterns of the animal.

[0071] Preferably, pixel-level matching can adopt an improved optical flow method, such as the Farneback optical flow algorithm, which can handle a large range of displacements and has a certain robustness to noise. When calculating the displacement vector, the grayscale value of each pixel can be differentially calculated to determine its displacement in the horizontal and vertical directions. For example, for each pixel, the grayscale difference in the horizontal direction in the next frame image is calculated. If the difference is greater than a certain threshold, it is considered that the pixel has been displaced in the horizontal direction. Similarly, the displacement in the vertical direction can be calculated. When generating a motion vector map, the displacement vector of each pixel can be normalized to facilitate subsequent analysis and visualization. For example, the length of the displacement vector can be normalized to a range of 0 to 1 while keeping its direction unchanged. The horizontal motion vector map and the vertical motion vector map generated in this way can more intuitively display the movement direction and speed distribution of the animal target, providing accurate data support for subsequent motion intensity analysis.

[0072] In some embodiments, calculating the motion intensity distribution includes:

[0073] According to the displacement of each pixel in the horizontal motion vector map and the movement speed threshold of the corresponding species, high activity area and low activity area are divided;

[0074] The same processing is performed on the longitudinal motion vector map, and the transverse and longitudinal activity areas are superimposed to generate a two-dimensional motion intensity distribution map.

[0075] It should be noted that the process of calculating the motion intensity distribution is to divide high-activity and low-activity areas based on the displacement of each pixel in the horizontal motion vector map, combined with the movement speed threshold of the corresponding species. The same process is performed on the vertical motion vector map, and the horizontal and vertical activity areas are superimposed to generate a two-dimensional motion intensity distribution map. The motion intensity distribution here refers to the activity intensity of the animal in different areas of the image. The activity areas are divided by displacement and speed threshold, which provides a more intuitive display of the animal's activities. The movement speed threshold is a baseline value set based on the average movement speed of the animal, which is used to distinguish between high-activity and low-activity areas.

[0076] Specifically, the displacement of each pixel in the horizontal motion vector map represents the intensity of motion of the pixel in the horizontal direction, and the displacement of each pixel in the vertical motion vector map represents the intensity of motion of the pixel in the vertical direction. The movement speed threshold is set according to the average movement speed of the animal. For example, for fast-moving animals, the threshold can be set higher, while for slow-moving animals, the threshold can be set lower. High activity areas refer to areas where the displacement exceeds the movement speed threshold, while low activity areas refer to areas where the displacement is lower than the threshold. By superimposing the horizontal and vertical activity areas, a two-dimensional motion intensity distribution map can be generated, which can intuitively show the distribution of animal activity intensity in different directions.

[0077] Preferably, the movement speed threshold can be determined by analyzing the movement data of a specific species. For example, by collecting the movement speed data of the species in different environments, calculating its average speed and standard deviation, and then setting the threshold to the average speed plus or minus one standard deviation. When calculating the motion intensity distribution, the displacement of each pixel can be normalized to facilitate comparison of the motion intensity in different directions. For example, the horizontal and vertical displacements are normalized to the range of 0 to 1 respectively, and then the activity areas are divided according to the comparison between the normalized displacements and the movement speed threshold. The generated two-dimensional motion intensity distribution map can use color coding to represent different activity intensities. For example, high activity areas are represented by red and low activity areas are represented by blue, which can more intuitively display the activity patterns of animals.

[0078] In some embodiments, the generation of the dynamic weight map includes:

[0079] Performing similarity matching on the motion intensity distribution map and typical activity templates in a species activity pattern database, and adjusting the weight coefficient of each region according to the matching result;

[0080] The weight coefficients are smoothed in time series to generate a dynamic weight map for spatiotemporal fusion.

[0081] It should be noted that the dynamic weight map is generated by similarity matching the motion intensity distribution map with typical activity templates in the species activity pattern database. The weight coefficients of each region are adjusted based on the matching results, and the weight coefficients are smoothed over time to generate a dynamic weight map for spatiotemporal fusion. The dynamic weight map here is a weight distribution map used to enhance key areas in image processing, and it can dynamically adjust weights based on the animal's activity patterns. Similarity matching refers to determining the adjustment method for weight coefficients by calculating the similarity between the motion intensity distribution map and the typical activity template. Time series smoothing is a processing method used to reduce the fluctuation of weight coefficients over time series to ensure weight stability.

[0082] Specifically, the motion intensity distribution map is generated by analyzing the horizontal and vertical motion vector maps, which reflects the activity intensity of animals in different areas. The species activity pattern database stores typical activity templates of various animals, which are obtained by analyzing a large amount of animal behavior data. The similarity matching process is achieved by calculating the similarity between the motion intensity distribution map and the typical activity template. The similarity can be measured by calculating the correlation coefficient or Euclidean distance between the two distributions. The weight coefficient is adjusted according to the similarity matching results. For example, if the activity pattern of a certain area is highly similar to the typical template, the weight coefficient of the area will be increased. Time series smoothing can be achieved by moving average method or exponential smoothing method to reduce the fluctuation of the weight coefficient in the time series and ensure its stability.

[0083] Preferably, similarity matching can be achieved by calculating the Pearson correlation coefficient between the motion intensity distribution map and the typical activity template. For example, for a specific species, its typical activity template shows a higher activity intensity in the central area, while the activity intensity in the edge area is lower. If the correlation coefficient between the motion intensity distribution map and the template is high, it means that the current animal's activity pattern is similar to the typical pattern of the species, and the weight coefficient of the corresponding area can be increased. Time series smoothing can use exponential smoothing, which can give higher weights to recent data while smoothing out short-term fluctuations. For example, the smoothing parameter can be set to 0.5, so that the update of the weight coefficient will be smoother, avoiding drastic fluctuations in the weight coefficient due to short-term activity changes. The dynamic weight map generated in this way can better reflect the activity pattern of the animal and provide a more accurate weight distribution for subsequent spatiotemporal fusion.

[0084] In some embodiments, the spatiotemporal fusion includes:

[0085] The animal target area of ​​the current frame is locally enhanced according to the dynamic weight map, and the background area is denoised using the multi-frame mean of historical frames;

[0086] The species classification labels are retained during the fusion process, and the motion trajectories of the same target in consecutive frames are connected into a complete path.

[0087] It should be noted that spatiotemporal fusion involves locally enhancing the target animal region in the current frame according to a dynamic weight map, applying multi-frame averaging of historical frames to the background region, retaining the species classification label during the fusion process, and connecting the motion trajectories of the same target in consecutive frames into a complete path. Spatiotemporal fusion is an image processing technique that combines information in both time and space to improve image quality and information content. The dynamic weight map is a weight distribution map dynamically generated based on the animal's activity patterns and is used to enhance the signal in the target region. Multi-frame averaging is a method for reducing noise by calculating the average value of multiple frames. The species classification label refers to the information about the animal species retained during image processing for subsequent analysis and identification.

[0088] Specifically, the dynamic weight map is dynamically generated based on the activity pattern of the animal. It contains the weight value of each pixel, which reflects the importance of the pixel in the image. During the spatiotemporal fusion process, the animal target area of ​​the current frame will be locally enhanced according to the dynamic weight map, which means that the pixel values ​​of the target area will be adjusted according to the weight value to highlight the target features. The background area is processed by multi-frame mean denoising, that is, the average value of the corresponding pixels in multiple frames is taken to reduce the interference of background noise. The species classification label is the key information retained in the image processing process. It identifies the type of animal in the image. These labels will not be lost during the fusion process to facilitate subsequent analysis and identification. The connection of motion trajectories is achieved by analyzing the position changes of the same target in consecutive frames, connecting these position points to form a complete motion path.

[0089] Preferably, the weight values ​​of the dynamic weight map can be adjusted according to the activity intensity and position of the animal. For example, if the activity intensity in a certain area is high, the weight value of the area can be set larger. When performing local enhancement, it can be achieved by multiplying the pixel value of the current frame by the corresponding weight value. For multi-frame mean denoising, a certain number of historical frames, such as 5 frames or 10 frames, can be selected to calculate the average value of each pixel to reduce background noise. When connecting motion trajectories, interpolation methods can be used to smooth the trajectory lines, such as linear interpolation or spline interpolation, to ensure the continuity and smoothness of the trajectory. In addition, in order to improve the accuracy of the trajectory, a distance threshold can be set. When the distance between two consecutive trajectory points exceeds the threshold, the trajectory is considered to be interrupted and needs to be reconnected. Through these refined steps, spatiotemporal fusion can process image data more efficiently and provide more accurate results for wildlife monitoring.

[0090] In some embodiments, the method further comprises a model adaptive updating step:

[0091] When an unidentified species is detected for a preset number of consecutive frames and the target area ratio exceeds a preset area threshold, the image acquisition device is automatically triggered to save the target area and adjacent frame images;

[0092] After performing background subtraction and histogram equalization preprocessing on the collected images, pseudo labels are generated through semi-supervised learning, and the species classification model is updated using an online knowledge distillation algorithm;

[0093] During incremental training, the feature extraction layer is frozen and only the classification layer parameters are updated. After training is completed, a model validation test is performed, and an updated version is deployed when the accuracy improvement exceeds the preset accuracy threshold.

[0094] It should be noted that the present invention relates to a model adaptive update mechanism. When an unidentified species is detected for a preset number of consecutive frames and the target area exceeds a preset area threshold, the image acquisition device is automatically triggered to save the target area and adjacent frame images. After performing background subtraction and histogram equalization preprocessing on the captured images, pseudo labels are generated through semi-supervised learning, and the species classification model is updated using an online knowledge distillation algorithm. During incremental training, the feature extraction layer is frozen, and only the classification layer parameters are updated. After training is completed, a model validation test is performed, and an updated version is deployed when the accuracy rate exceeds a preset precision threshold. Model adaptive updating here refers to the system's ability to automatically identify and adapt to new species. By saving image data of unidentified species, preprocessing and model updates are performed to improve the model's accuracy and adaptability. Background subtraction is an image processing technique used to remove background noise and highlight the target object. Histogram equalization is an image enhancement technique used to adjust the contrast of an image. Semi-supervised learning is a machine learning method that combines a small amount of labeled data with a large amount of unlabeled data for training. The online knowledge distillation algorithm is a model optimization technique that improves the performance and efficiency of a complex model by transferring knowledge from a simple model.

[0095] Specifically, the "preset number of consecutive frames" means that the system triggers an update mechanism when the number of frames in which unidentified species are detected reaches a set threshold. For example, this can be set to 10 consecutive frames. The target area ratio refers to the proportion of the area occupied by the unidentified species in the image. When it exceeds a set threshold, such as 10%, the species is considered worthy of further identification. Background difference removes background noise by calculating the difference between the current frame and a background model. The background model can be established by averaging the initial few frames. Histogram equalization adjusts the image's histogram distribution to achieve more uniform contrast. In semi-supervised learning, pseudo-labels are obtained by screening prediction results from unlabeled data, for example, selecting predictions with a confidence level above 0.8 as pseudo-labels. The online knowledge distillation algorithm improves the performance of the student model (a simpler model) by using the output of the teacher model (a complex model) as the training target. Incremental training involves training based on an existing model, updating only the parameters of the classification layer to reduce computational complexity. The preset accuracy threshold is the minimum improvement in accuracy after a model update, for example, it can be set to 2%.

[0096] Preferably, background difference can be achieved by calculating the pixel difference between the current frame and the background model. For example, when the pixel difference exceeds a certain threshold, such as 10 gray levels, the pixel is considered to belong to the target object. Histogram equalization can be achieved by calculating the cumulative histogram of the image and mapping it to a new grayscale range, for example, mapping the cumulative histogram to a grayscale range of 0 to 255. In semi-supervised learning, pseudo labels can be generated by setting a confidence threshold, such as 0.9, and selecting prediction results with confidence higher than the threshold as pseudo labels. In the online knowledge distillation algorithm, the teacher model can be a pre-trained deep convolutional neural network, and the student model can be a lightweight network. The student model is trained by minimizing the output difference between the teacher model and the student model. During incremental training, transfer learning technology can be used to freeze the feature extraction layer parameters and only train the classification layer, for example, using the Adam optimizer with a learning rate set to 0.001. Model validation testing can be done by evaluating the accuracy of the model on a validation set. When the accuracy improves beyond a preset accuracy threshold, such as 3%, the updated model is deployed into the monitoring system to improve the ability to identify new species.

[0097] In some embodiments, the data compression and transmission step includes:

[0098] Inter-frame compression coding is used for static background areas, and the compression rate is dynamically adjusted according to the wireless network signal strength;

[0099] When animal motion trajectory data is converted into vector format, each trajectory point contains timestamp, thermal radiation intensity, and motion direction data, and differential coding is used to compress the direction angle;

[0100] Prioritize uploading classification information containing preset protected species, trigger emergency transmission mode when abnormal movement patterns are detected, and improve the keyframe compression quality to above the preset threshold.

[0101] It should be noted that the data compression and transmission steps include inter-frame compression coding for static background areas, with the compression rate dynamically adjusted based on wireless network signal strength. When animal trajectory data is converted to vector format, each trajectory point includes a timestamp, thermal radiation intensity, and movement direction data, and differential coding is used to compress the directional angle. Priority is given to uploading classification information containing pre-set protected species. When abnormal movement patterns are detected, emergency transmission mode is triggered, improving the compression quality of key frames to above a preset threshold. Inter-frame compression coding is a video compression technology that reduces data volume by comparing the differences between consecutive frames. Wireless network signal strength refers to the quality of the wireless communication link, which affects the speed and reliability of data transmission. Vector format is a data representation method that uses geometric elements such as points and lines to describe movement trajectories. Differential coding is a data compression method that reduces data volume by storing the change in data rather than the original data. Emergency transmission mode is a mechanism that increases the priority of data transmission when an abnormality is detected.

[0102] Specifically, inter-frame compression coding reduces redundant data by encoding the differences between consecutive frames. For example, if the background area does not change in consecutive frames, only the data of the changed part is transmitted. The signal strength of the wireless network can be measured by parameters such as the signal strength indicator RSSI. When the signal strength is low, the compression rate can be reduced to reduce transmission errors. The trajectory points in the vector format contain timestamps, which record the time when the event occurred, the intensity of thermal radiation, and reflect the thermal characteristics and movement direction data of the animal, indicating the direction of movement of the animal. Differential coding compresses the direction angle. For example, if the direction angle of consecutive trajectory points changes little, only the change is transmitted. The emergency transmission mode is activated when an abnormal movement pattern is detected. For example, when the movement trajectory of the animal shows abnormal behavior, the system will improve the compression quality of the key frame to ensure the accurate transmission of important data.

[0103] Preferably, inter-frame compression coding can utilize standard video compression algorithms such as H.264, which effectively reduce data volume while maintaining image quality. Dynamic adjustment of wireless network signal strength can be achieved by monitoring the RSSI value. For example, when the RSSI value falls below -70dBm, the compression rate can be reduced from 50% to 30% to ensure data transmission reliability. Vector-format motion trajectory data can be stored by defining a structure, where the timestamp can be millisecond timestamps, the thermal radiation intensity can be a normalized value, and the motion direction data can be an angle value. Differential coding can compress directional angles by calculating the difference in directional angles between adjacent trajectory points. For example, if the directional angle change is less than 5 degrees, only the difference is transmitted. Emergency transmission mode can be triggered by setting an abnormal behavior threshold. For example, if the animal's movement speed suddenly changes by more than 20%, the system will increase the compression quality of key frames to above a preset threshold, such as 90%, to ensure that these key frames are transmitted with higher quality, thereby providing more accurate data for subsequent analysis.

[0104] In some embodiments, the abnormal behavior analysis includes:

[0105] Calculate the steering angles formed by adjacent trajectory points in the motion trajectory. If the cumulative steering angle exceeds an angle threshold within a preset time period and the number of occurrences of a single steering angle exceeding a sub-angle threshold reaches a preset number, it is determined to be wandering behavior;

[0106] Speed ​​mutation detection uses sliding window variance analysis to generate a poaching warning when the speed change exceeds the species average speed threshold within a preset time window and the frequency of occurrence reaches a preset frequency;

[0107] When analyzing multi-species trajectories, the degree of overlap of activity areas of different species is detected through spatial topology maps. When the proportion of overlapping area exceeds the area threshold and the duration exceeds the time threshold, the habitat conflict level is comprehensively determined in combination with multimodal sensor data.

[0108] It should be noted that abnormal behavior analysis involves calculating the turning angles formed by adjacent points in a trajectory. Wandering behavior is identified when the cumulative turning angle exceeds an angle threshold within a preset time period and the number of occurrences of a single turning angle exceeding a sub-angle threshold reaches a preset number. Speed ​​abrupt changes are detected using sliding window variance analysis. A poaching alert is generated when the speed change exceeds a species average speed threshold and the frequency of occurrence reaches a preset frequency within a preset time window. Multi-species trajectory analysis uses a spatial topology map to detect the overlap of activity areas of different species. When the proportion of overlapping area exceeds an area threshold and the duration exceeds a time threshold, the habitat conflict level is comprehensively determined using multimodal sensor data. Abnormal behavior analysis here refers to identifying possible abnormal behaviors, such as wandering or poaching, by analyzing animal trajectories. Turning angle refers to the angular change between adjacent points in an animal's trajectory. Sliding window variance analysis is a statistical method used to detect whether speed changes are abnormal. A spatial topology map is a map that represents the distribution of activity areas of different species. Multimodal sensor data refers to data from multiple sensors, such as sound and temperature, used for comprehensive analysis.

[0109] Specifically, turning angles are determined by calculating the angles between adjacent track points. For example, if the angle between two track points exceeds a certain threshold, such as 30 degrees, a significant turn is considered. Angle thresholds and sub-angle thresholds are set parameters to distinguish between normal and abnormal turns. For example, the angle threshold can be set to 180 degrees, and the sub-angle threshold can be set to 45 degrees. Preset time periods and preset number of times are used to determine wandering behavior. For example, if the cumulative turning angle exceeds 360 degrees and the number of single turning angles greater than 45 degrees exceeds 5 within 10 minutes, it is considered wandering behavior. In sliding window variance analysis, the sliding window size can be set based on the average speed of the species. For example, for faster animals, the sliding window can be set to 10 seconds; for slower animals, the sliding window can be set to 30 seconds. The species average speed threshold is set based on the normal movement speed of the animal. For example, if the average speed of a particular animal is 2 meters per second, the speed threshold can be set to 4 meters per second. The preset frequency is a parameter used to determine whether sudden changes in speed are frequent. For example, if the frequency of sudden changes in speed exceeds three per minute, a poaching alert is generated. In the spatial topology map, activity area overlap is determined by calculating the percentage of overlapping areas between the activity areas of different species. For example, if the overlap exceeds 50% and lasts for more than one hour, the level of habitat conflict is determined using a comprehensive analysis of multimodal sensor data.

[0110] Preferably, the turning angle can be calculated using the vector difference method, that is, calculating the angle between the direction vectors of adjacent trajectory points. For example, if the direction vectors of two trajectory points are x1, y1 and x2, y2, respectively, the turning angle can be calculated by calculating the angle between these two vectors. The angle threshold and sub-angle threshold can be adjusted based on the animal's behavioral characteristics. For example, for animals that frequently turn, the sub-angle threshold can be appropriately increased. In sliding window variance analysis, a moving average method can be used to calculate the average speed within the sliding window, and then the variance can be calculated to detect abnormal speed changes. For example, if the speed variance within the sliding window exceeds a certain threshold, such as 1 meter per second squared, the speed change is considered abnormal. When analyzing multi-species trajectories, a spatial topological map can be constructed using Geographic Information System (GIS) technology, and the degree of overlap of activity areas can be determined by calculating the ratio of the area of ​​the overlapping area to the total area. For example, if the activity areas of two species are A and B, and the overlapping area is C, the degree of overlap can be calculated by calculating the ratio of the area of ​​C to the sum of the areas of A and B. When combining multimodal sensor data, machine learning algorithms, such as decision trees or random forests, can be used to comprehensively analyze the data from different sensors, thereby more accurately determining the level of habitat conflict. For example, if a sound sensor detects an unusual sound and a temperature sensor detects a change in temperature, the level of habitat conflict can be increased.

[0111] The above-mentioned embodiments of the present invention have the following beneficial effects: the present invention can achieve a comprehensive improvement in wildlife monitoring efficiency. An infrared sensing trigger mechanism can intelligently control image acquisition timing, significantly reducing energy consumption while ensuring monitoring effectiveness. A lightweight network architecture combined with thermal radiation normalization processing can achieve efficient target detection on edge devices. Motion vector analysis technology can accurately capture animal movement characteristics, and dynamic weight map generation can optimize monitoring weights for different areas. A spatiotemporal fusion algorithm can improve image quality and reconstruct complete activity trajectories. An adaptive model update mechanism can continuously expand the system's recognition capabilities. An intelligent data compression strategy can dynamically adjust transmission parameters based on network conditions to ensure that critical data is transmitted first.

[0112] This invention enables multi-dimensional ecological monitoring. Infrared thermal imaging technology overcomes the time and environmental limitations of traditional monitoring; motion intensity distribution calculations enable quantitative analysis of animal behavioral characteristics; matching species activity databases improves the accuracy of behavioral identification; abnormal trajectory detection algorithms enable timely detection of potentially threatening behaviors; multi-species spatial analysis assesses habitat interactions; and semi-supervised learning mechanisms continuously optimize system performance. By integrating infrared sensing, intelligent analysis, and wireless transmission technologies, a complete intelligent closed-loop monitoring system can be constructed, providing all-weather, high-precision technical support for wildlife protection.

[0113] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, server, mobile phone, or tablet.

[0114] The above descriptions merely illustrate some preferred embodiments of the present invention and the underlying technical principles. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present invention is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A wildlife channel monitoring method based on infrared sensing and AI recognition, characterized in that: The steps include: When the infrared sensing module detects that the heat change in the channel exceeds a first preset threshold, it triggers high-resolution infrared image acquisition to generate an infrared image sequence of the wildlife channel; Performing animal target recognition on each frame of the infrared image sequence using an AI target detection model, extracting the motion area of ​​the animal target in each frame and generating a horizontal motion vector diagram and a vertical motion vector diagram relative to the previous frame; Calculate the motion intensity distribution of the animal target in each frame of the image based on the horizontal motion vector diagram and the vertical motion vector diagram, and generate a dynamic weight map in combination with the species activity pattern database; The dynamic weight map is spatially and temporally fused with the infrared images of the current frame and the historical frames, and the monitoring results with species classification and activity trajectory are output and uploaded to the data center via a wireless network; The generation of the dynamic weight graph includes: The motion intensity distribution map is matched with the typical activity templates in the species activity pattern database for similarity, and the weight coefficient of each area is adjusted according to the matching results; Perform time series smoothing on the weight coefficients to generate a dynamic weight map for spatiotemporal fusion; The space-time fusion includes: The animal target area of ​​the current frame is locally enhanced according to the dynamic weight map, and the background area is denoised using the multi-frame mean of historical frames; The species classification labels are retained during the fusion process, and the motion trajectories of the same target in consecutive frames are connected into a complete path; Abnormal behavior analysis includes: Calculate the steering angles formed by adjacent trajectory points in the motion trajectory. If the cumulative steering angle exceeds an angle threshold within a preset time period and the number of occurrences of a single steering angle exceeding a sub-angle threshold reaches a preset number, it is determined to be wandering behavior; Speed ​​mutation detection uses sliding window variance analysis to generate a poaching warning when the speed change exceeds the species average speed threshold within a preset time window and the frequency of occurrence reaches a preset frequency; When analyzing multi-species trajectories, the degree of overlap of activity areas of different species is detected through spatial topology maps. When the proportion of overlapping area exceeds the area threshold and the duration exceeds the time threshold, the habitat conflict level is comprehensively determined in combination with multimodal sensor data.

2. The method according to claim 1, characterized in that The triggering of high-resolution infrared image acquisition includes: Monitor pixel-level thermal changes in the infrared sensor module in real time. When the thermal change in the same area for a preset number of consecutive frames exceeds the ambient noise threshold, the infrared camera is activated to shoot at the highest resolution and switches to low-power mode when there is no activity.

3. The method according to claim 1, characterized in that The construction of the AI ​​target detection model includes: A lightweight convolutional neural network is used as the backbone network, and a thermal radiation intensity normalization layer is added to the infrared image input layer to linearly map the original thermal radiation value to a standard grayscale range. When loading a pre-trained species classification model through transfer learning, freeze the parameters of the feature extraction layer except the classification layer; When deploying edge computing devices, a model quantization engine is used to quantize the network at a low bit width to achieve real-time detection of a preset frame rate. The detection threshold is set to a preset confidence threshold and a preset intersection-to-union ratio threshold.

4. The method according to claim 1, characterized in that The generating of the horizontal motion vector map and the vertical motion vector map includes: Perform pixel-level matching on the motion areas of the same animal target in two adjacent frames of images, and calculate the horizontal and vertical displacement vectors based on the matching results; The displacement vector is mapped to the entire image according to the spatial position to generate a horizontal motion vector map and a vertical motion vector map respectively.

5. The method according to claim 1, characterized in that: The calculating of the motion intensity distribution includes: According to the displacement of each pixel in the horizontal motion vector map and the movement speed threshold of the corresponding species, high activity area and low activity area are divided; The same processing is performed on the longitudinal motion vector map, and the transverse and longitudinal activity areas are superimposed to generate a two-dimensional motion intensity distribution map.

6. The method according to claim 3, characterized in that: The method further comprises a model adaptive updating step: When an unidentified species is detected for a preset number of consecutive frames and the target area ratio exceeds a preset area threshold, the image acquisition device is automatically triggered to save the target area and adjacent frame images; After performing background subtraction and histogram equalization preprocessing on the collected images, pseudo labels are generated through semi-supervised learning, and the species classification model is updated using an online knowledge distillation algorithm; During incremental training, the feature extraction layer is frozen and only the classification layer parameters are updated. After training is completed, a model validation test is performed, and an updated version is deployed when the accuracy improvement exceeds the preset accuracy threshold.

7. The method according to claim 1, characterized in that: The data compression and transmission step comprises: Inter-frame compression coding is used for static background areas, and the compression rate is dynamically adjusted according to the wireless network signal strength; When animal motion trajectory data is converted into vector format, each trajectory point contains timestamp, thermal radiation intensity, and motion direction data, and differential coding is used to compress the direction angle; Prioritize uploading classification information containing preset protected species, trigger emergency transmission mode when abnormal movement patterns are detected, and improve the keyframe compression quality to above the preset threshold.

Images