An edge AI-based local image recognition terminal device

Abstract

The application relates to the technical field of artificial intelligence and embedded systems, and discloses a local image recognition terminal device based on edge AI, wherein a perception layer supports multi-modal data cooperative collection, and environment data such as images, temperature and humidity, gas and sound are synchronously acquired; an edge computing layer is loaded with a lightweight AI model, local inference and multi-modal data fusion are completed, and the cloud is not needed to be relied on; a power management module has multi-mode power supply and hibernation wake-up capabilities, and low-power operation is realized; and a communication and control module supports multi-protocol remote interaction and OTA upgrading. The application realizes local AI inference and multi-modal data fusion processing through the edge computing layer, does not need to rely on a cloud server, and completes data transmission and processing locally, so that the problem of insufficient real-time performance caused by network dependence in a traditional architecture is effectively solved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and embedded systems technology, specifically referring to a local image recognition terminal device based on edge AI. Background Technology

[0002] Traditional image recognition systems generally adopt a "terminal acquisition-cloud processing" architecture, where the terminal device is only responsible for acquiring image data, and then all data is uploaded to a cloud server for centralized inference and analysis. This architecture has revealed many insurmountable limitations in practical applications. It relies entirely on a stable network connection for data transmission and processing command feedback. In remote areas, regions with weak network infrastructure, or scenarios with network congestion, data transmission latency is prone to excessive delays or even communication interruptions, causing recognition tasks to fail to complete in a timely manner. For scenarios with extremely high response speed requirements, such as traffic violation recognition and factory equipment anomaly monitoring, the millisecond-level feedback requirements of cloud processing are difficult to meet, severely impacting system availability.

[0003] The uploading of large amounts of sensitive image data (such as facial information, behavioral trajectories, and images of core industrial scenes) to the cloud via the network not only increases the risk of data leakage during transmission and storage but also poses significant threats to user privacy and industry data security. Most existing terminal devices rely solely on cameras to acquire visual information. Single-modal data is easily affected by factors such as changes in lighting, object occlusion, and environmental interference, leading to a decrease in image recognition accuracy. Currently, there are few smart terminals on the market that support multi-sensor collaborative operation, and those that exist are mostly concentrated in the high-end market, resulting in high costs and hindering large-scale application. Most image recognition terminals require 24-hour continuous operation, especially in continuous monitoring scenarios, where device power consumption is high. Existing terminals lack effective low-power management mechanisms, resulting in short battery life and frequent charging, which not only increases user costs but also limits their deployment in outdoor or unreliable power supply environments.

[0004] To address the aforementioned shortcomings of existing technologies, this invention proposes a local image recognition terminal device based on edge AI to solve the problems of cloud dependence, privacy and security, limited sensing capabilities, and excessive power consumption. Summary of the Invention

[0005] In response to the above situation, this paper proposes a local image recognition terminal device based on edge AI to address the shortcomings of existing image recognition systems, such as strong cloud dependence, high data privacy and security risks, limited environmental perception capabilities, and poor power consumption control.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a local image recognition terminal device based on edge AI, comprising: It includes a perception layer, which is used to acquire images and multi-dimensional environmental data and supports multi-modal data collaborative acquisition; An edge computing layer is used to perform AI inference and multimodal data fusion processing locally. The power management module is used to realize low power supply and energy consumption control, and has multi-mode power supply adaptation and sleep wake-up capabilities. A communication and control module, which is used for remote interaction, device configuration, and data transmission; And an alarm and interaction module, which is used for graded alarms for abnormal events and local visual operation interaction; The perception layer is communicatively connected to the edge computing layer to transmit the collected multimodal data to the edge computing layer in real time for local inference and fusion processing; The edge computing layer is communicatively connected to the power management module, the communication and control module, and the alarm and interaction module. The connection with the power management module is used to dynamically adjust the power supply mode according to the task load to optimize energy consumption. The connection with the communication and control module is used to transmit inference results, alarm information, and receive remote configuration and upgrade commands. The connection with the alarm and interaction module is used to trigger hierarchical alarm actions and respond to local operation commands.

[0007] As a further description of the above technical solution: The perception layer includes an image acquisition module and a multi-sensor interface module; The image acquisition module is a high-definition camera module used to acquire stable image data; The multi-sensor interface module integrates multiple standard communication interfaces such as I2C, SPI, and UART, and can selectively connect to at least two of the following: infrared sensor, temperature and humidity sensor, microphone array, and gas sensor. The infrared sensor is used for human movement detection, the temperature and humidity sensor is used for real-time monitoring of environmental temperature and humidity parameters, the microphone array is used for collecting environmental sound information, and the gas sensor is used for detecting air quality or concentration of harmful gases.

[0008] As a further description of the above technical solution: The edge computing layer includes edge AI computing units and an operating framework; The edge AI computing unit uses a high-performance embedded AI chip, which integrates a multi-core ARM architecture CPU, a parallel computing GPU, and an NPU specifically for neural network inference optimization, and is equipped with memory and storage units to meet the local AI inference computing needs.

[0009] As a further description of the above technical solution: The image recognition model running on the edge computing layer is a lightweight neural network model, with an input resolution of no less than 320×320 and an inference speed of no less than 15 FPS. The model training and optimization methods include: pre-training based on public datasets, continuously optimizing the model's generalization ability through local online incremental learning, and using knowledge distillation and pruning techniques to achieve lightweight model compression, ensuring efficient operation at the edge.

[0010] As a further description of the above technical solution: The edge computing layer is configured with a multi-sensor data fusion algorithm, which includes a data synchronization strategy and a fusion strategy. The data synchronization strategy employs timestamp alignment and Kalman filtering techniques to achieve precise time-series synchronization of image data and multi-sensor data. The fusion strategy includes at least one of early fusion, mid-term fusion, and late-term fusion. Early fusion involves stitching the image with the original sensor data to form a multimodal input. Mid-term fusion involves extracting features of each modality and then performing feature-level fusion. Late-term fusion involves weighted voting decision-making on the recognition results of each modality. The algorithm also processes time-series environmental data through an LSTM network and employs an attention mechanism to increase the weight ratio of key features in the fusion process.

[0011] As a further description of the above technical solution: The power management module adopts dynamic voltage and frequency adjustment technology and has two core working modes: standby mode and task identification mode. The power management module supports a sleep-wake mechanism, which can wake up the device by triggering an external sensor or a timer. The power supply method includes at least one of the following: main power supply, auxiliary power supply, and external power supply. The main power supply uses a lithium battery, the auxiliary power supply is a combination of a solar panel and an MPPT charging controller, and the external power supply supports USB Type-C interface or DC power adapter.

[0012] As a further description of the above technical solution: The communication and control module integrates a multi-protocol wireless communication module and supports wireless communication protocols.

[0013] As a further description of the above technical solution: The alarm and interaction module includes a local alarm unit, a remote notification unit, and a control panel; The local alarm unit consists of a buzzer and red, yellow, and green LED indicator lights; the remote notification unit can push alarm information to users via SMS, email, and mobile APP; the control panel is a touch screen or physical buttons, supporting local parameter settings and device status queries. The module has a built-in alarm rule engine, which sets the image recognition confidence threshold and the abnormal range of sensor data, and adopts a hierarchical alarm strategy of primary local sound and light prompts, secondary remote message notifications, and tertiary cloud platform manual intervention.

[0014] As a further description of the above technical solution: The lightweight model of the edge computing layer is deployed in the TensorFlowLite+FlatBuffers format. The edge computing layer optimizes the image preprocessing process through OpenCV. In the process of multimodal data fusion, a unified data format is used to achieve cross-modal data alignment. Furthermore, the ONNX model is deployed at the edge to support joint inference of multimodal data.

[0015] As a further description of the above technical solution: The power management module also includes a low-power collaboration mechanism, in which the MCU is responsible for continuously monitoring the sensor status and waking up the edge AI computing unit only when a valid event is detected. The edge computing layer adopts multi-task scheduling. During the multi-sensor data synchronization process, Kalman filtering technology is used to filter out environmental interference noise, improve the accuracy of data fusion, and ensure the reliability of event judgment.

[0016] The beneficial effects achieved by the present invention using the above structure are as follows: (1) In this invention, local AI inference and multimodal data fusion processing are achieved through an edge computing layer, without relying on a cloud server. Data transmission and processing are completed locally on the terminal, and the response latency is controlled at the millisecond level. This effectively solves the problem of insufficient real-time performance caused by network dependence in traditional architectures, and is especially suitable for scenarios with high requirements for response speed. At the same time, image data and multi-dimensional environmental data are collected, processed and stored locally without uploading to the cloud, avoiding the risk of leakage of sensitive data during transmission and cloud storage, which is conducive to protecting user privacy and industry data security.

[0017] (2) In this invention, a multi-sensor data fusion algorithm is used to achieve comprehensive analysis of cross-modal information, effectively resisting environmental interference such as light and occlusion, significantly improving the accuracy of event judgment in complex scenarios, and reducing false alarm rate and false negative rate. The power management module adopts dynamic voltage and frequency adjustment technology, sleep wake-up mechanism and low power consumption coordination mechanism to achieve intelligent energy consumption control. The standby power consumption is <0.5W and the power consumption in the recognition task mode is <5W. At the same time, it supports multiple power supply methods such as lithium battery, solar energy and external power supply, adapting to various deployment scenarios such as the field and no stable power supply, extending the device's battery life and reducing the cost of use. Attached Figure Description

[0018] Figure 1This is a schematic diagram of the structure of the local image recognition terminal device based on edge AI proposed in this invention; Figure 2 This is a flowchart illustrating the workflow of the local image recognition terminal device based on edge AI proposed in this invention. Figure 3 This is a schematic diagram of the working structure of the local image recognition terminal device based on edge AI proposed in this invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0021] To address the issues of strong cloud dependency, high data privacy and security risks, limited environmental perception capabilities, and poor power consumption control in image recognition systems, this application presents a local image recognition terminal device based on edge AI, such as... Figures 1-3 As shown, it includes a perception layer, which is used to acquire images and multi-dimensional environmental data, and supports multi-modal data collaborative acquisition; Edge computing layer: The edge computing layer is used to complete AI inference and multimodal data fusion processing locally; The power management module is used to achieve low-power supply and energy consumption control, and has multi-mode power supply adaptation and sleep / wake-up capabilities. The communication and control module is used for remote interaction, device configuration, and data transmission. And an alarm and interaction module, which is used for graded alarms for abnormal events and local visual operation interaction; The perception layer communicates with the edge computing layer to transmit the collected multimodal data to the edge computing layer in real time for local inference and fusion processing; The edge computing layer is connected to the power management module, the communication and control module, and the alarm and interaction module. The connection with the power management module is used to dynamically adjust the power supply mode according to the task load to optimize energy consumption. The connection with the communication and control module is used to transmit inference results, alarm information and receive remote configuration and upgrade commands. The connection with the alarm and interaction module is used to trigger hierarchical alarm actions and respond to local operation commands.

[0022] In this way, cross-modal information comprehensive analysis can be achieved through multi-sensor data fusion algorithms, effectively resisting environmental interference such as lighting and occlusion, significantly improving the accuracy of event judgment in complex scenarios, and reducing false alarm and false negative rates. The power management module adopts dynamic voltage and frequency adjustment technology, sleep-wake mechanism, and low-power collaborative mechanism to achieve intelligent energy consumption control. Standby power consumption is <0.5W, and power consumption in recognition task mode is <5W. It also supports multiple power supply methods such as lithium battery, solar power, and external power supply, adapting to various deployment scenarios such as outdoor and unstable power supply, extending the device's battery life and reducing usage costs.

[0023] The perception layer includes an image acquisition module and a multi-sensor interface module. The image acquisition module is a high-definition camera module with autofocus, night vision enhancement, and HDR dynamic range extension capabilities, which can acquire clear and stable image data. The multi-sensor interface module integrates multiple standard communication interfaces such as I2C, SPI, and UART, and can selectively connect to at least two of the following: infrared sensors, temperature and humidity sensors, microphone arrays, and gas sensors. Among them, the infrared sensor is used for human movement detection, the temperature and humidity sensor is used for real-time monitoring of environmental temperature and humidity parameters, the microphone array is used to collect environmental sound information, and the gas sensor is used to detect air quality or the concentration of harmful gases.

[0024] The edge computing layer includes edge AI computing units and a runtime framework. The edge AI computing units use high-performance embedded AI chips that integrate a multi-core ARM architecture CPU, a GPU with parallel computing acceleration capabilities, and an NPU specifically designed for neural network inference optimization. They are also equipped with memory and storage units that meet the needs of local AI inference operations. The runtime framework is compatible with mainstream embedded AI inference frameworks such as TensorFlow Lite, ONNXRuntime, and PyTorchMobile, and can support the rapid deployment and efficient operation of lightweight models.

[0025] The image recognition model running on the edge computing layer is a lightweight neural network model with an input resolution of no less than 320×320 and an inference speed of no less than 15 FPS. After lightweight processing, the model size does not exceed 10MB. The model training and optimization methods include: pre-training based on public datasets, continuously optimizing the model's generalization ability through local online incremental learning, and using knowledge distillation and pruning techniques to achieve lightweight compression of the model, ensuring efficient operation at the edge.

[0026] The edge computing layer is configured with a multi-sensor data fusion algorithm, which includes a data synchronization strategy and a fusion strategy. The data synchronization strategy uses timestamp alignment and Kalman filtering to achieve accurate temporal synchronization between image data and multi-sensor data. The fusion strategy includes at least one of early fusion, mid-term fusion, and late fusion. Early fusion involves concatenating the image and the original sensor data to form a multimodal input. Mid-term fusion involves extracting features from each modality and then performing feature-level fusion. Late fusion involves weighted voting decisions based on the recognition results of each modality. The algorithm also processes temporal environmental data through an LSTM network and uses an attention mechanism to increase the weight of key features in the fusion process, thereby improving the accuracy of event judgment.

[0027] The power management module adopts Dynamic Voltage and Frequency Scaling (DVFS) technology and has two core operating modes: standby mode and identification task mode. The power consumption in standby mode is <0.5W, and the power consumption in identification task mode is <5W. The power management module supports a sleep-wake mechanism, which can be triggered by external sensors or timers to wake up the device. The power supply method includes at least one of the following: main power supply, auxiliary power supply, and external power supply. The main power supply uses a lithium battery, the auxiliary power supply is a combination of a solar panel and an MPPT charging controller, and the external power supply supports USB Type-C interface or DC power adapter to adapt to different deployment scenarios.

[0028] The communication and control module integrates a multi-protocol wireless communication module, supporting various wireless communication protocols including Wi-Fi 802.11b / g / n, Bluetooth 5.0, NB-IoT, and LoRa, adapting to both short-range and wide-area network communication needs. The module is compatible with MQTT, HTTP, and CoAP remote communication protocols, enabling remote configuration of device parameters, real-time push of alarm information, and feedback on device operating status. The module supports OTA upgrades, employing differential upgrade technology to reduce update package size, adapting to model and system upgrade needs in remote areas or weak network environments.

[0029] The alarm and interaction module includes a local alarm unit, a remote notification unit, and a control panel. The local alarm unit consists of a buzzer and red, yellow, and green LED indicators, providing on-site audible and visual alarm prompts. The remote notification unit can push alarm information to users via SMS, email, and mobile APP. The control panel is a touch screen or physical buttons, supporting local parameter settings and device status queries. The module has a built-in alarm rule engine, which allows for custom settings of image recognition confidence thresholds and sensor data anomaly ranges. It adopts a hierarchical alarm strategy of primary local audible and visual prompts, secondary remote message notifications, and tertiary cloud platform manual intervention. It can also dynamically adjust alarm sensitivity based on environmental information such as time, location, and weather to reduce false alarm rates.

[0030] The lightweight model of the edge computing layer is deployed in the TensorFlowLite+FlatBuffers format, with a model loading time of <50ms. The edge computing layer optimizes the image preprocessing process through OpenCV and uses hardware acceleration technology to improve image processing efficiency. In the process of multimodal data fusion, a unified data format is used to achieve cross-modal data alignment, and the ONNX model is deployed at the edge to support multimodal data joint inference, further improving processing efficiency and recognition accuracy.

[0031] The power management module also includes a low-power coordination mechanism, where the MCU is responsible for continuously monitoring the sensor status and only waking up the edge AI computing unit when a valid event is detected, reducing unnecessary power consumption. The edge computing layer uses a real-time operating system to implement multi-task scheduling, ensuring the real-time performance of processes such as data acquisition, inference calculation, and alarm response. During the multi-sensor data synchronization process, Kalman filtering technology is used to filter out environmental interference noise, improve the accuracy of data fusion, and ensure the reliability of event judgment.

[0032] To better understand the working process of a local image recognition terminal device based on edge AI according to an embodiment of this application, refer to... Figures 1-3 The following is a specific embodiment: After the device is started, the power management module enters standby mode by default, and the MCU continuously monitors the sensor status. Users can configure parameters locally through the control panel of the alarm and interaction module, or remotely through the MQTT / HTTP / CoAP protocol of the communication and control module. The configuration parameters include image recognition confidence threshold, sensor data anomaly range, alarm sensitivity, wireless communication mode, etc. At the same time, the edge computing layer loads a lightweight neural network model, completes the inference framework initialization, and waits for data input.

[0033] When the multi-sensor interface module of the perception layer detects human movement through the infrared sensor or is triggered by a timer, the power management module switches the device to the recognition task mode and wakes up the image acquisition module and all connected sensors. The image acquisition module activates autofocus, night vision enhancement, and HDR dynamic range extension functions to acquire real-time image data, ensuring clear and stable images. The multi-sensor interface module synchronously acquires multi-dimensional environmental data such as temperature and humidity, ambient sound, air quality, or concentration of harmful gases through the IC / SPI / UART interface. The acquired multi-modal data is transmitted to the edge computing layer in real time through the communication connection and is preprocessed: the image data is cropped, normalized, and color space converted (e.g., RGB→YUV), and the sensor data is filtered, denoised, and standardized.

[0034] The edge computing layer receives preprocessed multimodal data and uses a runtime framework (TensorFlowLite / ONNXRuntime / PyTorchMobile) to call a lightweight neural network model for local inference. The image recognition model performs target detection and classification on the image data, outputting the target category, location, and confidence level. The model inference speed is no less than FPS to ensure real-time performance. The multi-sensor data fusion algorithm module starts working: the data synchronization unit uses timestamp alignment and Kalman filtering technology to achieve accurate time synchronization between image data and sensor data, while filtering out environmental interference noise; the fusion strategy unit selects early fusion, mid-term fusion, or late fusion methods according to the scenario requirements to comprehensively analyze the multimodal data; the time-series data processing unit processes time-series data such as temperature and humidity change trends through an LSTM network and uses an attention mechanism to enhance the weights of key features, finally outputting the event judgment result.

[0035] The edge computing layer compares the event judgment result with a preset threshold. If no anomaly is triggered, the device maintains the recognition task mode and continues to collect and process data. If no valid event is triggered for more than a preset time, the power management module automatically switches to standby mode to reduce energy consumption. If an anomaly is triggered, the edge computing layer sends an alarm command to the alarm and interaction module to activate a tiered alarm: for a primary alarm, the buzzer of the local alarm unit is activated, and the corresponding red / yellow / green LED indicator lights up to provide on-site audio and visual prompts; for a secondary alarm, the remote notification unit pushes alarm information to the user via SMS, email, and mobile APP, including details such as event type, occurrence time, and environmental parameters; for a tertiary alarm, if the user does not respond within a preset time, the communication and control module links the alarm information to the cloud platform to request manual intervention.

[0036] The communication and control module transmits device operating status, inference results, and alarm information to user terminals or cloud platforms in real time, while also receiving remote configuration commands to dynamically adjust device parameters. When model upgrades or system updates are required, the communication and control module obtains the update package through OTA differential upgrade functionality and transmits it to the edge computing layer to complete the upgrade. The upgrade process does not affect the core functions of the device. Throughout the entire operation, the low-power collaborative mechanism of the power management module continues to operate: the MCU monitors the sensor status and only wakes up the edge AI computing unit when a valid event is detected; the edge computing layer uses a real-time operating system to implement multi-task scheduling and dynamically adjusts the GPU / NPU operating frequency according to the task load to further optimize energy consumption; when the device is deployed in outdoor scenarios, it can switch to a solar + lithium battery power supply mode, and the MPPT charging controller ensures efficient charging of the solar panels, guaranteeing continuous operation of the device.

[0037] This invention achieves core functions such as local intelligent identification, low-power operation, multimodal data fusion, hierarchical alarm and remote operation and maintenance through the collaborative work of various modules. It effectively solves many defects of existing technologies, has significant technical advancement and practical value, can be widely applied to various edge computing scenarios, and has good prospects for industrialization and promotion.

[0038] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0039] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0041] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A local image recognition terminal device based on edge AI, characterized in that, include: It includes a perception layer, which is used to acquire images and multi-dimensional environmental data and supports multi-modal data collaborative acquisition; An edge computing layer is used to perform AI inference and multimodal data fusion processing locally. The power management module is used to realize low power supply and energy consumption control, and has multi-mode power supply adaptation and sleep wake-up capabilities. A communication and control module, which is used for remote interaction, device configuration, and data transmission; And an alarm and interaction module, which is used for graded alarms for abnormal events and local visual operation interaction; The perception layer is communicatively connected to the edge computing layer to transmit the collected multimodal data to the edge computing layer in real time for local inference and fusion processing; The edge computing layer is communicatively connected to the power management module, the communication and control module, and the alarm and interaction module. The connection with the power management module is used to dynamically adjust the power supply mode according to the task load to optimize energy consumption. The connection with the communication and control module is used to transmit inference results, alarm information, and receive remote configuration and upgrade commands. The connection with the alarm and interaction module is used to trigger hierarchical alarm actions and respond to local operation commands.

2. The local image recognition terminal device based on edge AI according to claim 1, characterized in that, The perception layer includes an image acquisition module and a multi-sensor interface module; The image acquisition module is a high-definition camera module used to acquire stable image data; The multi-sensor interface module integrates multiple standard communication interfaces such as I2C, SPI, and UART, and can selectively connect to at least two of the following: infrared sensor, temperature and humidity sensor, microphone array, and gas sensor. The infrared sensor is used for human movement detection, the temperature and humidity sensor is used for real-time monitoring of environmental temperature and humidity parameters, the microphone array is used for collecting environmental sound information, and the gas sensor is used for detecting air quality or concentration of harmful gases.

3. The local image recognition terminal device based on edge AI according to claim 2, characterized in that, The edge computing layer includes edge AI computing units and an operating framework; The edge AI computing unit uses a high-performance embedded AI chip, which integrates a multi-core ARM architecture CPU, a parallel computing GPU, and an NPU specifically for neural network inference optimization, and is equipped with memory and storage units to meet the local AI inference computing needs.

4. The local image recognition terminal device based on edge AI according to claim 3, characterized in that, The image recognition model running on the edge computing layer is a lightweight neural network model, with an input resolution of no less than 320×320 and an inference speed of no less than 15 FPS. The model training and optimization methods include: pre-training based on public datasets, continuously optimizing the model's generalization ability through local online incremental learning, and using knowledge distillation and pruning techniques to achieve lightweight model compression, ensuring efficient operation at the edge.

5. A local image recognition terminal device based on edge AI according to claim 4, characterized in that, The edge computing layer is configured with a multi-sensor data fusion algorithm, which includes a data synchronization strategy and a fusion strategy. The data synchronization strategy employs timestamp alignment and Kalman filtering techniques to achieve precise time-series synchronization of image data and multi-sensor data. The fusion strategy includes at least one of early fusion, mid-term fusion, and late-term fusion. Early fusion involves stitching the image with the original sensor data to form a multimodal input. Mid-term fusion involves extracting features of each modality and then performing feature-level fusion. Late-term fusion involves weighted voting decision-making on the recognition results of each modality. The algorithm also processes time-series environmental data through an LSTM network and employs an attention mechanism to increase the weight ratio of key features in the fusion process.

6. A local image recognition terminal device based on edge AI according to claim 5, characterized in that, The power management module adopts dynamic voltage and frequency adjustment technology and has two core working modes: standby mode and task identification mode. The power management module supports a sleep-wake mechanism, which can wake up the device by triggering an external sensor or a timer. The power supply method includes at least one of the following: main power supply, auxiliary power supply, and external power supply. The main power supply uses a lithium battery, the auxiliary power supply is a combination of a solar panel and an MPPT charging controller, and the external power supply supports USB Type-C interface or DC power adapter.

7. A local image recognition terminal device based on edge AI according to claim 6, characterized in that, The communication and control module integrates a multi-protocol wireless communication module and supports wireless communication protocols.

8. A local image recognition terminal device based on edge AI according to claim 7, characterized in that, The alarm and interaction module includes a local alarm unit, a remote notification unit, and a control panel; The local alarm unit consists of a buzzer and red, yellow, and green LED indicator lights; the remote notification unit can push alarm information to users via SMS, email, and mobile APP; the control panel is a touch screen or physical buttons, supporting local parameter settings and device status queries. The module has a built-in alarm rule engine, which sets the image recognition confidence threshold and the abnormal range of sensor data, and adopts a hierarchical alarm strategy of primary local sound and light prompts, secondary remote message notifications, and tertiary cloud platform manual intervention.

9. A local image recognition terminal device based on edge AI according to claim 8, characterized in that, The lightweight model of the edge computing layer is deployed in the TensorFlowLite+FlatBuffers format. The edge computing layer optimizes the image preprocessing process through OpenCV. In the process of multimodal data fusion, a unified data format is used to achieve cross-modal data alignment. Furthermore, the ONNX model is deployed at the edge to support joint inference of multimodal data.

10. A local image recognition terminal device based on edge AI according to claim 9, characterized in that, The power management module also includes a low-power collaboration mechanism, in which the MCU is responsible for continuously monitoring the sensor status and waking up the edge AI computing unit only when a valid event is detected. The edge computing layer adopts multi-task scheduling. During the multi-sensor data synchronization process, Kalman filtering technology is used to filter out environmental interference noise, improve the accuracy of data fusion, and ensure the reliability of event judgment.

Images