Wild animal monitoring data optimization method and system based on edge calculation

Abstract

The invention discloses a wild animal monitoring data optimization method and system based on edge computing, and relates to the technical field of wild animal monitoring, and the method comprises the steps: collecting image or video data of a target region in real time through a sensor connected with an edge computing device, and carrying out the data preprocessing; performing real-time analysis on the preprocessed image or video data by using a lightweight deep learning model deployed on the edge computing device, and identifying a target species or an abnormal behavior; based on the identification result, dynamically screening valid data including the target species image or video, and deleting invalid data; and carrying out compression coding on the screened effective data, and storing the data in a local storage unit or transmitting the data to a cloud end through a low-power-consumption network. The problems that in an existing method, no network area data return cost is high, the period is long, the storage and manual processing burden is heavy due to the fact that the invalid data proportion is high, and existing edge equipment lacks dynamic compression and self-adaptive power management capacity are solved.

Description

Technical Field

[0001] The present invention relates to the field of wildlife monitoring technology, and in particular to a wildlife monitoring data optimization method and system based on edge computing. Background Art

[0002] The existing wildlife monitoring system has the following shortcomings:

[0003] Reliance on manual data transmission: Infrared cameras require regular manual recycling of memory cards, which dramatically increases workload when data volumes are large (e.g., a single device generates 1TB of data per month).

[0004] High proportion of invalid data: Traditional devices cannot identify invalid data in real time (such as blank screens triggered by mistake), resulting in waste of storage and transmission resources;

[0005] Difficulties in deployment in areas without network: Remote areas cannot transmit data in real time via 4G / 5G and must rely on satellite communications, which is costly. Summary of the Invention

[0006] In response to the above-mentioned deficiencies in the prior art, the present invention provides a wildlife monitoring data optimization method and system based on edge computing, which solves the problems in the existing methods such as high cost and long cycle of data backhaul in areas without network, high proportion of invalid data resulting in heavy burden of storage and manual processing, and lack of dynamic compression and adaptive power management capabilities of existing edge devices.

[0007] In order to achieve the above-mentioned purpose of the invention, the technical solution adopted by the present invention is: a wildlife monitoring data optimization method based on edge computing, comprising the following steps:

[0008] S1: Sensors connected to edge computing devices collect image or video data of the target area in real time and perform data preprocessing;

[0009] S2: Analyze the pre-processed image or video data in real time using a lightweight deep learning model deployed on the edge computing device to identify target species or abnormal behaviors;

[0010] S3: Dynamically screen valid data including images or videos of target species based on the recognition results and delete invalid data;

[0011] S4: Compress and encode the filtered valid data and store it in a local storage unit or transmit it to the cloud via a low-power network.

[0012] Furthermore, the edge computing device in S1 is configured with an adaptive power management module, and the adaptive power management module includes:

[0013] Solar power supply device;

[0014] Sleep-wake-up mechanism: The sleep-wake-up mechanism turns off the camera and computing unit when no animal activity is detected, and wakes up the device when motion is detected.

[0015] Furthermore, the data preprocessing in S1 includes bilinear interpolation, Gaussian filtering, normalization and affine transformation.

[0016] Furthermore, the lightweight deep learning model in S2 includes:

[0017] An object detection network based on EfficientDet is used to locate the position of target species in the image;

[0018] A classification network that adds a dual-stream feature extraction network and a feature interaction fusion network to the ResNet network is used for fine-grained recognition of target species.

[0019] The dual-stream feature extraction network includes a global feature stream and a local feature stream. The global feature stream extracts the overall features of the image through the Transformer module, and the local feature stream is implemented through the attention mechanism and local key area focusing to identify the specific representation parts of the current animal that distinguish it from other species.

[0020] Furthermore, the output F of the global feature flow g for:

[0021] F g =Transformer(I)

[0022] Among them, Transformer(·) is the Transformer module, I is the primary image feature information obtained by the ResNet network;

[0023] The output F1 of the local feature flow is:

[0024] M=σ(Conv(F g ))

[0025] F1=M*F g

[0026] Among them, M is the weight of spatial attention, σ is the Sigmoid function, Conv(·) is the convolution process, and * represents element-by-element multiplication.

[0027] Furthermore, the output F of the feature interaction fusion network pooled for:

[0028] F pooled =Sign(F g *F1)*log(|F g *F1|+ε)

[0029] Where Sign(·) is the sign function, and ε is a very small positive number whose value is 1e-5.

[0030] Furthermore, deleting invalid data in S3 includes the following steps:

[0031] S31: Eliminate false trigger images caused by environmental interference through background subtraction algorithm;

[0032] S32: Filter data irrelevant to the target species based on the preset species, and only retain valid data with a confidence level higher than the set threshold.

[0033] Furthermore, the compressed encoding of the filtered valid data in S4 includes the following sub-steps:

[0034] S41: Dynamically adjust compression parameters, including frame rate and resolution, based on the behavioral characteristics of the target species;

[0035] S42: Use high compression rate encoding for static background areas in the image, and use lossless compression that retains detail information for areas where animals are active.

[0036] The present invention also adopts a technical solution: a system for optimizing wildlife monitoring data based on edge computing, comprising:

[0037] Edge computing devices: integrated NPU chips for data collection, local recognition, and dynamic screening;

[0038] Data management module: includes local storage unit, low-power communication module and adaptive power management unit;

[0039] Cloud management platform: used to receive valid data, generate ecological analysis reports, remotely update the algorithm model of edge devices, and monitor device status.

[0040] The beneficial effects of the present invention are:

[0041] (1) Data volume reduction: invalid data deletion rate > 90%, and the storage requirement of a single device is reduced to at least 10GB / month;

[0042] (2) Reduced labor costs: Data collection previously required monthly manual maintenance, but now only requires quarterly maintenance;

[0043] (3) Improved real-time performance: key events (such as poaching) trigger satellite backhaul with a delay of less than 1 minute;

[0044] (4) Improved model recognition accuracy: The accuracy of fine-grained classification of target species reached 98.2%;

[0045] (5) Energy efficiency optimization: Edge device power consumption is reduced by 60% (combination of dynamic sleep and solar power supply);

[0046] (6) The present invention can be extended to fields such as bird monitoring, agricultural pest monitoring, and invasive species prevention and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a wildlife monitoring data optimization method based on edge computing in the present invention.

[0048] Figure 2 Flowchart for lightweight deep learning model processing. DETAILED DESCRIPTION

[0049] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0050] Example 1, as Figure 1 As shown, a wildlife monitoring data optimization method based on edge computing includes the following steps:

[0051] S1: Sensors connected to edge computing devices collect image or video data of the target area in real time and perform data preprocessing;

[0052] S2: Analyze the pre-processed image or video data in real time using a lightweight deep learning model deployed on the edge computing device to identify target species or abnormal behaviors;

[0053] S3: Dynamically screen valid data including images or videos of target species based on the recognition results and delete invalid data;

[0054] S4: Compress and encode the filtered valid data and store it in a local storage unit or transmit it to the cloud via a low-power network.

[0055] The edge computing device in S1 is configured with an adaptive power management module, and the adaptive power management module includes:

[0056] Solar power supply device;

[0057] It can automatically switch between solar and battery power according to the ambient light, giving priority to solar power when there is sufficient light, and switching to battery power at night or when there is insufficient light.

[0058] Sleep-wake-up mechanism: The sleep-wake-up mechanism turns off the camera and computing unit when no animal activity is detected, and wakes up the device when motion is detected.

[0059] The data preprocessing in S1 includes bilinear interpolation, Gaussian filtering, normalization and affine transformation.

[0060] The local preprocessing module primarily utilizes machine learning theories and methods, applying preprocessing techniques such as bilinear interpolation, Gaussian filtering, normalization, and affine transformation to images containing wildlife input. This module performs preliminary filtering and processing of input data, providing structured information for the feature extraction model.

[0061] The lightweight deep learning model in S2 includes:

[0062] An object detection network based on EfficientDet is used to locate the position of target species in the image;

[0063] A classification network that adds a dual-stream feature extraction network and a feature interaction fusion network to the ResNet network is used for fine-grained recognition of target species.

[0064] The dual-stream feature extraction network includes a global feature stream and a local feature stream. The global feature stream extracts the overall features of the image through the Transformer module, and the local feature stream is implemented through the attention mechanism and local key area focusing to identify the specific representation parts of the current animal that distinguish it from other species.

[0065] The lightweight deep learning model used to identify target species or abnormal behaviors consists of two parts, such as Figure 2 As shown, one is to use the EfficientDet network to detect the position of animals in the picture, and the other is to add a dual-stream feature extraction and feature interaction fusion module on the basis of the ResNet network to achieve refined classification of wild animals and improve the ability to distinguish similar species. Among them, the dual-stream feature extraction network is further composed of a global feature stream and a local feature stream. The global feature stream further analyzes the primary image features obtained by the ResNet network by adding a Transformer module to obtain a higher level of overall feature information; the local feature stream is realized through the attention mechanism and local key area focusing to identify the specific characterization parts of a certain animal that are different from other species. In addition, the present invention compresses the model to less than 10MB through knowledge distillation and quantization technology, making it suitable for low-computing power infrared camera equipment.

[0066] The output F of the global feature flow g for:

[0067] F g =Transformer(I)

[0068] Among them, Transformer(·) is the Transformer module, I is the primary image feature information obtained by the ResNet network;

[0069] The output F1 of the local feature flow is:

[0070] M=σ(Conv(F g ))

[0071] F1=M*F g

[0072] Among them, M is the weight of spatial attention, σ is the Sigmoid function, Conv(·) is the convolution process, and * represents element-by-element multiplication.

[0073] The output F of the feature interaction fusion network pooled for:

[0074] F pooled =Sign(F g *F1)*log(|F g *F1|+ε)

[0075] Where Sign(·) is the sign function, and ε is a very small positive number whose value is 1e-5.

[0076] The core of the network is bilinear pooling, which interacts global and local features to improve fine-grained classification capabilities.

[0077] Deleting invalid data in S3 includes the following steps:

[0078] S31: Eliminate false trigger images caused by environmental interference through background subtraction algorithm;

[0079] S32: Filter data irrelevant to the target species based on the preset species, and only retain valid data with a confidence level higher than the set threshold.

[0080] In this embodiment, valid data is determined based on a confidence threshold (e.g., >0.8); a background difference algorithm is used to eliminate false trigger images caused by environmental interference such as swaying leaves; and only images or video clips containing preset target species are retained.

[0081] The present invention synchronizes timestamp-based images with environmental data (temperature, light, etc.) to solve the problem of spatiotemporal inconsistency in multi-source data; and uses environmental data to assist in judging animal behavior (such as the probability of animal appearance at a specific temperature).

[0082] The S4 compresses and encodes the filtered valid data, including the following steps:

[0083] S41: Dynamically adjust compression parameters, including frame rate and resolution, based on the behavioral characteristics of the target species;

[0084] S42: Use high compression rate encoding for static background areas in the image, and use lossless compression that retains detail information for areas where animals are active.

[0085] When storing compressed valid data in a local storage unit or transmitting it to the cloud via a low-power network, phased transmission can be adopted, that is, low-priority data (such as regular species activities) is transmitted through the LoRa self-organizing network, and emergency events (such as poaching) trigger satellite backhaul.

[0086] In one embodiment of the present invention, taking the Siberian tiger monitoring scenario as an example, the specific implementation is as follows:

[0087] (1) Hardware deployment and initialization

[0088] 1) Deploy edge devices: Install edge computing devices in Siberian tiger habitats (such as the Changbai Mountain Nature Reserve) and configure:

[0089] Sensor: Infrared camera (trigger sensitivity set to "animal heat + motion detection");

[0090] Communication module: LoRa module (transmission distance 10 kilometers, battery life 3 years);

[0091] Power supply system: solar panel (20W) + lithium battery (12V / 10Ah), supports dynamic sleep (turns off the camera at night when there are no animals, and power consumption is reduced to 0.1W).

[0092] 2) Initialize the model and parameters: Remotely download a lightweight model (<10MB) via the cloud platform:

[0093] EfficientDet-Lite: an object detection model that supports detection of over 100 species, with a Siberian tiger detection confidence threshold of 0.8.

[0094] Two-stream ResNet classification model: Based on the improvement of ResNet-18, it adds Transformer global stream and attention local stream, and the accuracy of species fine-grained classification reaches 98%.

[0095] (2) Data collection and local processing

[0096] 1) Environmental triggering and data capture

[0097] Motion detection: When the infrared camera detects a thermal signal (animal body temperature > 25°C) or a motion trigger (pixel change rate > 5%), it starts recording (30 seconds / clip).

[0098] Data preprocessing: Perform the following on the captured images:

[0099] Bilinear interpolation: unify the resolution to 1080P;

[0100] Gaussian filter: remove noise (σ=1.5);

[0101] Normalization: Pixel values ​​are scaled to [0,1].

[0102] 2) Object Detection (EfficientDet)

[0103] Input: Preprocessed image (3 channels, resolution 1080P).

[0104] Output: Bounding Box of the Siberian tiger and confidence score (e.g. 0.92).

[0105] Processing time: ≤50ms (NPU acceleration).

[0106] 3) Species Classification (Two-Stream ResNet)

[0107] Global Feature Flow (Transformer):

[0108] Input the primary features of EfficientDet (C=256, H×W=14×14);

[0109] The Transformer encoder (2 layers, 4 heads) extracts the global feature Fg to capture the overall characteristics of the Siberian tiger, such as body shape and posture.

[0110] Local feature flow (attention mechanism): Generates attention weights to highlight local features such as the Siberian tiger's stripes and tail. Local feature extraction preserves key regional features.

[0111] Feature interaction fusion (bilinear pooling): The fully connected layer outputs species probabilities (Siberian tiger: 0.98, other species: 0.02).

[0112] (3) Dynamic screening and data optimization

[0113] 1) Invalid data deletion

[0114] Background differential screening:

[0115] Perform differential analysis on three consecutive frames of images. If the change rate is less than 2% (such as slight shaking of leaves), it is determined to be invalid data and deleted.

[0116] Experimental data: The false trigger image deletion rate is >85%.

[0117] Species-independent data filtering:

[0118] Only segments with confidence > 0.8 and species identified as Siberian tiger were retained, and data on other species (such as deer and fox) were deleted.

[0119] 2) Dynamic Compression

[0120] Adaptive compression strategy:

[0121] Animal activity area: lossless compression (JPEG2000 quality factor 100%);

[0122] Background area: high compression ratio (JPEG2000 quality factor 80%);

[0123] Frame rate adjustment: reduced to 1fps with a static background, increased to 15fps when the animal is moving.

[0124] Experimental data: After compression, the storage space of a single 30-second video is reduced from 150MB to 18MB, and the total storage requirement is reduced to 10GB / month (originally 1TB / month).

[0125] (4) Data transmission and cloud processing

[0126] 1) Low-power network transmission

[0127] The edge device sends the compressed data (e.g., video clips of Siberian tigers) to the relay node via LoRa.

[0128] Transmission priority: regular data (LoRa), emergency events (such as poaching sounds) trigger satellite backhaul (delay < 1 minute).

[0129] 2) Cloud processing

[0130] Ecological report generation:

[0131] Generate heat maps based on the time and location of Siberian tiger appearances to assist in habitat protection planning;

[0132] Experimental data: After using this system in a protected area, the accuracy of identifying the activity areas of Siberian tigers increased by 30%.

[0133] Model iteration:

[0134] The cloud collects new species data (such as newly added snow leopard monitoring) and updates the edge device model through knowledge distillation.

[0135] (5) Power management and adaptive control

[0136] 1) Dynamic sleep mechanism

[0137] Trigger conditions:

[0138] No motion detection for more than 10 minutes → enter sleep mode (power consumption 0.1W);

[0139] Detects animal heat signals → wakes up the camera and NPU within 1 second.

[0140] Experimental data: The average annual power consumption of edge devices is reduced by 60%, and solar power supply can meet all-weather operation.

[0141] 2) Environmental Adaptation

[0142] Lighting control:

[0143] When the light intensity is greater than 500 lux, solar charging is preferred;

[0144] Switch to battery power at night and the infrared camera automatically activates night vision mode.

[0145] Based on this embodiment, the following experimental data are obtained:

[0146] Model performance:

[0147] EfficientDet-Lite: achieves an average precision (AP) of 58.2% on the COCO wildlife dataset and an inference speed of 15 fps.

[0148] Two-stream ResNet classification: On the iWildCam dataset containing more than 100 species, fine-grained classification accuracy is 98.2% (compared to ResNet-50: 92.1%).

[0149] System efficiency:

[0150] Data volume reduction: 90% of invalid data is deleted, and the monthly storage requirement of a single device is reduced from 1TB to 10GB;

[0151] Power consumption optimization: Dynamic sleep reduces the average annual power consumption from 120W to 48W;

[0152] Transmission cost: LoRa transmission cost is only 1 / 10 of satellite communication.

[0153] By deploying 100 devices in the Changbai Mountain Nature Reserve and running them for 6 consecutive months:

[0154] The accuracy rate of Siberian tiger monitoring is 97%;

[0155] The false trigger rate was reduced from 35% to 5%;

[0156] The equipment maintenance cycle has been extended from once a month to once a quarter.

[0157] Example 2, a system for optimizing wildlife monitoring data based on edge computing, comprising:

[0158] Edge computing devices: integrated NPU chips for data collection, local recognition, and dynamic screening;

[0159] The hardware layer uses an infrared camera (resolution ≥ 1080P) as the visual sensor, which has night vision capabilities and adapts to varying lighting conditions. A wide-angle lens design expands the monitoring range. The NPU chip is optimized for lightweight model inference, increasing processing speed and reducing latency.

[0160] Data management module: includes local storage unit, low-power communication module and adaptive power management unit;

[0161] Cloud management platform: used to receive valid data, generate ecological analysis reports, remotely update the algorithm model of edge devices, and monitor device status.

[0162] The overall system architecture of the present invention includes hardware layer, data layer, algorithm layer and application layer. Among them, the hardware layer: edge devices deployed at monitoring nodes (such as ARM chips, NPU acceleration modules), integrating high-sensitivity microphone arrays and wide-angle cameras; data layer: real-time acquisition of image streams and local preprocessing (image enhancement); algorithm layer realizes refined classification of wild animals; application layer stores results locally or transmits key data to the central platform through low-power communication protocols (such as LoRa).

[0163] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the invention.

Claims

1. A wildlife monitoring data optimization method based on edge computing, characterized in that: The following steps are involved: S1: Sensors connected to edge computing devices collect image or video data of the target area in real time and perform data preprocessing; S2: Analyze the pre-processed image or video data in real time using a lightweight deep learning model deployed on the edge computing device to identify target species or abnormal behaviors; S3: Dynamically screen valid data including images or videos of target species based on the recognition results and delete invalid data; S4: Compress and encode the filtered valid data and store it in a local storage unit or transmit it to the cloud via a low-power network.

2. The wildlife monitoring data optimization method based on edge computing according to claim 1 is characterized in that: The edge computing device in S1 is configured with an adaptive power management module, and the adaptive power management module includes: Solar power supply device; Sleep-wake-up mechanism: The sleep-wake-up mechanism turns off the camera and computing unit when no animal activity is detected, and wakes up the device when motion is detected.

3. The wildlife monitoring data optimization method based on edge computing according to claim 1 is characterized in that: The data preprocessing in S1 includes bilinear interpolation, Gaussian filtering, normalization and affine transformation.

4. The wildlife monitoring data optimization method based on edge computing according to claim 1 is characterized in that: The lightweight deep learning model in S2 includes: An object detection network based on EfficientDet is used to locate the position of target species in the image; A classification network that adds a dual-stream feature extraction network and a feature interaction fusion network to the ResNet network is used for fine-grained recognition of target species. The dual-stream feature extraction network includes a global feature stream and a local feature stream. The global feature stream extracts the overall features of the image through the Transformer module, and the local feature stream is implemented through the attention mechanism and local key area focusing to identify the specific representation parts of the current animal that distinguish it from other species.

5. The wildlife monitoring data optimization method based on edge computing according to claim 4 is characterized in that: The output F of the global feature flow g for: F g =Transformer(I) Among them, Transformer(·) is the Transformer module, I is the primary image feature information obtained by the ResNet network; The output F1 of the local feature flow is: M=σ(Conv(F g )) F1=M*F g Among them, M is the weight of spatial attention, σ is the Sigmoid function, Conv(·) is the convolution process, and * represents element-by-element multiplication.

6. The wildlife monitoring data optimization method based on edge computing according to claim 5 is characterized in that: The output F of the feature interaction fusion network pooled for: F pooled =Sign(F g *F1)*log(|F g *F1|+ε) Where Sign(·) is the sign function, and ε is a very small positive number whose value is 1e-5.

7. The wildlife monitoring data optimization method based on edge computing according to claim 1 is characterized in that: Deleting invalid data in S3 includes the following steps: S31: Eliminate false trigger images caused by environmental interference through background subtraction algorithm; S32: Filter data irrelevant to the target species based on the preset species, and only retain valid data with a confidence level higher than the set threshold.

8. The method for optimizing wildlife monitoring data based on edge computing according to claim 1, wherein: In said S4, the filtered valid data is compressed and encoded. The following steps are included: S41: Dynamically adjust compression parameters, including frame rate and resolution, based on the behavioral characteristics of the target species; S42: Use high compression rate encoding for static background areas in the image, and use lossless compression that retains detail information for areas where animals are active.

9. A system for optimizing wildlife monitoring data based on edge computing according to any one of claims 1 to 8, characterized in that: include: Edge computing devices: integrated NPU chips for data collection, local recognition, and dynamic screening; Data management module: includes local storage unit, low-power communication module and adaptive power management unit; Cloud management platform: used to receive valid data, generate ecological analysis reports, remotely update the algorithm model of edge devices, and monitor device status.

Images