Edge distributed low latency intelligent cluster system

Abstract

The application discloses an edge distributed low-delay intelligent cluster system, belongs to the technical field of electric digital data processing, and solves the technical problems of real-time networking, low-delay computing and unstable network self-adaption of the edge computing cluster in a complex environment; the system comprises a master node, multiple slave nodes and a wireless Mesh communication unit, the master node and the slave nodes form a star distributed architecture through Mesh communication, and are both configured with an edge intelligent computing platform comprising an MLU220 SoC chip; the master node is internally provided with a TCP Cluster, an environment sensing, a link control and a node cooperation module, the slave node is internally provided with a target detection, identification and decision module, all the modules work cooperatively, and through dynamic load balancing, resolution and model switching of bandwidth self-adaption and node state closed-loop optimization, low-delay intelligent computing under an unstable network is realized, and the system is applicable to high-dynamic and high-reliability demand scenes such as unmanned aerial vehicle clusters and battlefield edge networks.

Description

Technical Field

[0001] This invention belongs to the field of electronic digital data processing technology, specifically relating to an edge-distributed low-latency intelligent cluster system. Background Technology

[0002] Edge computing deploys computing power near the data source, reducing transmission latency and cloud pressure, and is a key support for intelligent sensing and real-time decision-making. In complex environments such as battlefields and emergency rescue, edge clusters must meet core requirements such as strong real-time networking, low-latency computing, and adaptive capabilities to unstable networks. Existing technologies have the following shortcomings: edge nodes mostly use general-purpose processors, resulting in insufficient AI inference computing power and poor energy efficiency; traditional wireless communication lacks self-organization, self-relay, and self-healing capabilities, and has weak anti-interference and resilience; it lacks a bandwidth-task-model collaborative adaptive mechanism, leading to a sharp increase in latency during network fluctuations; and it has not achieved integrated collaboration between master-slave node dynamic sensing, link control, and computing power scheduling.

[0003] Therefore, there is an urgent need for a system that can solve the problems of low-latency computing and unstable network adaptation in edge clusters under complex environments. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes an edge-distributed low-latency intelligent cluster system. This system addresses the technical shortcomings of existing edge computing clusters in complex environments, such as insufficient AI inference computing power, weak communication anti-interference capabilities, lack of bandwidth-task-model adaptive mechanisms, and poor master-slave node collaboration. It designs a star-shaped distributed edge intelligent cluster architecture and a dynamic scheduling strategy for multi-module collaboration, thereby achieving low-latency intelligent computing under unstable network conditions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An edge-distributed low-latency intelligent cluster system includes a master node, multiple slave nodes, and a wireless mesh communication unit. The master node and slave nodes establish a communication connection through the wireless mesh communication unit, forming a star-shaped distributed cluster architecture. Both the master node and slave nodes are configured with an edge intelligent computing platform. The edge intelligent computing platform includes an AI computing unit, a power module, an RJ45 network port, a UART serial port, a 32GeMMCFLASH, and an LPDDR4 storage module. The master node has a built-in TCPCluster module, an environment awareness module, a link control module, and a node collaboration module. The slave nodes have a built-in target detection module, a target recognition module, and a target decision-making module. The TCP Cluster module includes a TCP cluster manager thread and several TCP send threads, which are synchronized through a message queue. The TCP cluster manager thread listens for TCP connection requests from slave nodes, selects TCP send threads according to the load balancing strategy, and dynamically creates or destroys threads. The TCP send threads are responsible for bidirectional data transmission with the corresponding slave nodes. The load balancing strategy is based on dynamic task allocation according to the load status of slave nodes, and the specific processing steps include: S1B, Calculate the load factor of the slave node. The calculation formula is as follows: ,in, This refers to the load factor of the slave node; Let m be the length of the task queue of the m-th slave node; This represents the maximum length of the queue, with a value of 10. This represents the percentage of computing power used by the m-th slave node. This represents the total computing power of the slave nodes; m is the slave node number. S2B, task allocation strategy: The master node obtains information from each slave node in real time. ,choose The smallest slave node is assigned the computation task; if all slave nodes... When necessary, the model will automatically be downgraded to reduce the computational load of a single task and ensure low-latency operation of the cluster. The link control module includes a network bandwidth awareness submodule and a task data adaptive allocation submodule. The network bandwidth awareness submodule queries the current Mesh network's communication bandwidth and transmission rate in real time using AT commands from the microhard PX2 module. The task data adaptive allocation submodule dynamically matches a deep learning model based on the network bandwidth awareness results and the number of online slave nodes, and adaptively adjusts the resolution of the image data to be transmitted, thereby achieving dynamic matching of transmission link bandwidth pressure. The specific processing includes: S1C, Bandwidth Acquisition: Commands are sent via the microhard PX2 module of the wireless mesh communication unit to obtain the real-time network transmission rate. The average bandwidth is calculated using a moving average filtering algorithm. The calculation formula is as follows: ,in, Average bandwidth; is the real-time network transmission rate for the k-th data acquisition; N is the number of data acquisitions. This is the index for the number of data collections; k ranges from 0 to 4. S2C, bandwidth tiering and binding to corresponding image resolutions and learning models: when At Mbps, this corresponds to an 800×600 resolution and the VGG16-Small model; When 15 Mbps < At Mbps, this corresponds to a 600×512 resolution and the VGG16-Middle model; when At Mbps, it corresponds to a 224×224 resolution and the VGG16-Large model; S3C, Model Switching Decision: Calculated using the delayed prediction model formula, the formula is as follows: ,in, is the end-to-end delay prediction value for a single frame image; W is the image width; H is the image height; This refers to the total computing power of the MLU220 SoC chip. when When the resolution is reduced and the model is switched, the slave nodes are notified to preload the target model. The environmental perception module collects computing load, execution latency, and bandwidth fluctuations, and calculates the state change rate using the following formula: ,in, The rate of change of state; The real-time value of the parameter at the current time t: These are the historical parameter values ​​for the previous 10 monitoring periods; When the rate of change of any one of the parameters—computing load, execution latency, or bandwidth fluctuation—satisfies |ΔX|>0.1, feedback is sent to the node collaboration module and link control module to optimize task allocation and model switching. The collaborative working relationship among the modules is as follows: the wireless Mesh communication unit collects network bandwidth and task data and transmits it to the link control module and the environment perception module; the TCPCluster module manages node connections and thread scheduling, and the node collaboration module performs computation task allocation; the AI ​​computing unit completes parallel inference computation and returns the results; the environment perception module monitors node load and network status in real time, and optimizes task allocation and model switching strategies in reverse to achieve low-latency intelligent computing in unstable network environments.

[0006] Preferably, the AI ​​computing unit uses the MLU220 SoC chip, integrating a quad-core ARM Cortex-A55 processor and an NPU deep learning processor, and runs on the Ubuntu 16.04 operating system and Neuware SDK software stack. It is used for lightweight deep learning model parallel inference, and the processing flow includes: S1A, Model Loading and Initialization: Load the lightweight deep learning model and resolve the model parallelism through the CNRT runtime library. , The value can be 1 or 4; data parallelism ,satisfy Initialize three independent streams: data copying, model inference, and result return. S2A, Inference Calculation: The received image data is processed using formulas... Normalization, in which, The image data before normalization. The image data is normalized. Parallel inference is performed based on the MLU220SoC chip core allocation strategy. The formula for calculating the inference time is as follows: , , ,in, The reasoning process takes time; The time taken to push data onto the stack; To calculate the time consumed; This refers to the time spent popping data from the stack. The data volume is represented by FLOPs, and the floating-point operation volume is represented by FLOPs. The inference time is input as an execution delay parameter to the environment perception module for node status monitoring and closed-loop optimization, and at the same time provides a computational basis for the delay prediction model of the link control module. S3A. Post-processing of results: The non-maximum suppression algorithm is used to filter duplicate recognition results. The calculation formula is as follows: ,in, This is a non-maximum suppression algorithm; To retrieve the index corresponding to the maximum value; Let be the confidence level of the i-th detection box; For intersection, union, and comparison; For the i-th detection box; For the j-th detection box; For all j and j not equal to i; i and All of these are detection box numbers.

[0007] Preferably, the wireless Mesh communication unit adopts a microhard PX2 module, which supports 802.11b / g / n high-speed wireless communication mode and four networking modes: AP / Client / Repeater / Mesh. It provides dual gigabit Ethernet ports, an RS232 serial port, and a USB 3.0 interface. The MLU220 SoC chip and the microhard PX2 module are interconnected through gigabit or 100 Mbps Ethernet ports and serial ports. The microhard PX2 modules achieve self-organization, self-relay, and self-healing long-distance wireless communication between nodes through Mesh networking.

[0008] Preferably, the LPDDR4 storage module is an external X32 LPDDR4X 3733Mbps chip, and the MLU220 SoC chip and the LPDDR4X memory chip are connected in a point-to-point direct topology.

[0009] Preferably, the node collaboration module includes an Image Handle submodule and a Web Controller submodule; the Image Handle submodule is used to complete multi-source data preprocessing and payload interface adaptation; the Web Controller submodule is used to parse the Note field in the network packet, start the inference pipeline, manage batch processing, and fill in the batch size when the transaction ends.

[0010] By adopting the above technical solution, the present invention has the following beneficial effects: 1. The AI ​​computing unit of this invention adopts the MLU220 SoC chip, which integrates a quad-core ARM Cortex A55 processor and an NPU deep learning processor, providing a total computing power of 16 TOPS@INT8 with a typical power consumption of only 8 watts and an energy efficiency ratio of 1.2 TOPS / W. This solves the problems of insufficient AI inference computing power and poor energy efficiency caused by the use of general-purpose processors in traditional edge nodes. At the same time, through the optimized configuration of model parallelism and data parallelism and the asynchronous execution of three independent streams, it realizes efficient parallel inference of lightweight deep learning models, with a collaborative recognition rate of 1000 frames / minute for 800×600 images and a single-node target recognition latency of ≤50ms, meeting the real-time requirements of intelligent perception at the edge.

[0011] 2. The wireless Mesh communication unit of this invention adopts a microhard PX2 module, which supports four networking modes: AP / Client / Repeater / Mesh. The modules achieve long-distance wireless communication between master and slave nodes through Mesh networking. It has the communication characteristics of self-organization, self-relay, and self-healing. A single group can support up to 16 nodes online at the same time. Even when nodes move, there is network interference, or some nodes are damaged, communication can still be maintained stably, which greatly improves the anti-interference and anti-destruction of the cluster and is suitable for highly dynamic application scenarios such as battlefields and drone swarms.

[0012] 3. This invention designs an integrated dynamic scheduling system. Through the load balancing strategy of the TCP Cluster module, it achieves precise allocation of computing tasks based on the load coefficient of slave nodes. When overloaded, it automatically triggers model degradation to avoid increased latency caused by single-node overload. Through the bandwidth awareness and adaptive allocation of the link control module, it achieves precise binding between network bandwidth and image resolution and deep learning models. Combined with the latency prediction model, it realizes active switching between resolution and model, eliminating transmission and computing latency caused by network fluctuations. Through the node status monitoring and closed-loop optimization of the environment awareness module, it captures sudden changes in computing load, execution latency, and bandwidth fluctuations in real time, and reversely optimizes task allocation and model switching strategies to ensure that the cluster can maintain low-latency deterministic computing under different network bandwidths and different node load conditions.

[0013] 4. This invention adopts a star-shaped distributed cluster architecture with global scheduling of the master node and distributed inference of the slave nodes. The functional modules of the master node and the detection / identification / decision modules of the slave nodes perform their respective functions and work together. It supports the collaborative computing power of at least 5 nodes and can be expanded to build a self-organizing cluster of no less than 10 computing nodes. At the same time, the edge intelligent computing platform integrates common interfaces such as RJ45 network port and UART serial port. The Image Handle submodule provides scalable payload interface adaptation, which is compatible with multi-source data processing such as image, radar and infrared, which greatly improves the hardware scalability and business adaptability of the system. The batch processing size padding mechanism of the WebController submodule avoids data blocking problems, ensures the continuous and efficient operation of the inference pipeline, and further improves the robustness of the cluster operation. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall system framework of the present invention; Figure 2 This is a schematic diagram of the main node of the present invention; Figure 3 This is a schematic diagram of the slave node framework of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] like Figures 1 to 3 As shown, an edge-distributed low-latency intelligent cluster system is characterized by comprising a master node, multiple slave nodes, and a wireless mesh communication unit; the master node and slave nodes establish a communication connection through the wireless mesh communication unit to form a star-shaped distributed cluster architecture; both the master node and slave nodes are configured with an edge intelligent computing platform; the edge intelligent computing platform includes an AI computing unit, a power module, an RJ45 network port, a UART serial port, a 32GeMMCFLASH, and an LPDDR4 storage module; the master node has a built-in TCPCluster module, an environment perception module, a link control module, and a node collaboration module; the slave nodes have a built-in target detection module, a target recognition module, and a target decision module. The TCP Cluster module includes a TCP cluster manager thread and several TCP send threads, which are synchronized through a message queue. The TCP cluster manager thread listens for TCP connection requests from slave nodes, selects TCP send threads according to the load balancing strategy, and dynamically creates or destroys threads. The TCP send threads are responsible for bidirectional data transmission with the corresponding slave nodes. The load balancing strategy is based on dynamic task allocation according to the load status of slave nodes, and the specific processing steps include: S1B, Calculate the load factor of the slave node. The calculation formula is as follows: ,in, This refers to the load factor of the slave node; Let m be the length of the task queue of the m-th slave node; This represents the maximum length of the queue, with a value of 10. This represents the percentage of computing power used by the m-th slave node. This represents the total computing power of the slave nodes; m is the slave node number. S2B, task allocation strategy: The master node obtains information from each slave node in real time. ,choose The smallest slave node is assigned the computation task; if all slave nodes... When necessary, the model will automatically be downgraded to reduce the computational load of a single task and ensure low-latency operation of the cluster. The link control module includes a network bandwidth awareness submodule and a task data adaptive allocation submodule. The network bandwidth awareness submodule queries the current Mesh network's communication bandwidth and transmission rate in real time using AT commands from the microhard PX2 module. The task data adaptive allocation submodule dynamically matches a deep learning model based on the network bandwidth awareness results and the number of online slave nodes, and adaptively adjusts the resolution of the image data to be transmitted, thereby achieving dynamic matching of transmission link bandwidth pressure. The specific processing includes: S1C, Bandwidth Acquisition: Commands are sent via the microhard PX2 module of the wireless mesh communication unit to obtain the real-time network transmission rate. The average bandwidth is calculated using a moving average filtering algorithm. The calculation formula is as follows: ,in, Average bandwidth; is the real-time network transmission rate for the k-th data acquisition; N is the number of data acquisitions. This is the index for the number of data collections; k ranges from 0 to 4. S2C, bandwidth tiering and binding to corresponding image resolutions and learning models: when At Mbps, this corresponds to an 800×600 resolution and the VGG16-Small model; When 15 Mbps < At Mbps, this corresponds to a 600×512 resolution and the VGG16-Middle model; when At Mbps, it corresponds to a 224×224 resolution and the VGG16-Large model; S3C, Model Switching Decision: Calculated using the delayed prediction model formula, the formula is as follows: ,in, is the end-to-end delay prediction value for a single frame image; W is the image width; H is the image height; This refers to the total computing power of the MLU220 SoC chip. when When the resolution is reduced and the model is switched, the slave nodes are notified to preload the target model. The environmental perception module collects computing load, execution latency, and bandwidth fluctuations, and calculates the state change rate using the following formula: ,in, The rate of change of state; The real-time value of the parameter at the current time t: These are the historical parameter values ​​for the previous 10 monitoring periods; When the rate of change of any one of the parameters—computing load, execution latency, or bandwidth fluctuation—satisfies |ΔX|>0.1, feedback is sent to the node collaboration module and link control module to optimize task allocation and model switching. The collaborative working relationship among the modules is as follows: the wireless Mesh communication unit collects network bandwidth and task data and transmits it to the link control module and the environment perception module; the TCPCluster module manages node connections and thread scheduling, and the node collaboration module performs computation task allocation; the AI ​​computing unit completes parallel inference computation and returns the results; the environment perception module monitors node load and network status in real time, and optimizes task allocation and model switching strategies in reverse to achieve low-latency intelligent computing in unstable network environments.

[0017] The AI ​​computing unit uses the MLU220 SoC chip, integrating a quad-core ARM Cortex-A55 processor and an NPU deep learning processor. It runs on the Ubuntu 16.04 operating system and the Neuware SDK software stack, and is used for lightweight deep learning model parallel inference. The processing flow includes: S1A, Model Loading and Initialization: Load the lightweight deep learning model and resolve the model parallelism through the CNRT runtime library. , The value can be 1 or 4; data parallelism ,satisfy Initialize three independent streams: data copying, model inference, and result return. S2A, Inference Calculation: The received image data is processed using formulas... Normalization, in which, The image data before normalization. The image data is normalized. Parallel inference is performed based on the MLU220SoC chip core allocation strategy. The formula for calculating the inference time is as follows: , , ,in, The reasoning process takes time; The time taken to push data onto the stack; To calculate the time consumed; This refers to the time spent popping data from the stack. The data volume is represented by FLOPs, and the floating-point operation volume is represented by FLOPs. The inference time is input as an execution delay parameter to the environment perception module for node status monitoring and closed-loop optimization, and at the same time provides a computational basis for the delay prediction model of the link control module. S3A. Post-processing of results: The non-maximum suppression algorithm is used to filter duplicate recognition results. The calculation formula is as follows: ,in, This is a non-maximum suppression algorithm; To retrieve the index corresponding to the maximum value; Let be the confidence level of the i-th detection box; For intersection, union, and comparison; For the i-th detection box; For the j-th detection box; For all j and j not equal to i; i and All of these are detection box numbers.

[0018] The wireless mesh communication unit uses a microhard PX2 module, which supports 802.11b / g / n high-speed wireless communication mode and four networking modes: AP / Client / Repeater / Mesh. It provides dual gigabit Ethernet ports, an RS232 serial port, and a USB 3.0 interface. The MLU220 SoC chip and the microhard PX2 module are interconnected through gigabit or 100-megabit Ethernet ports and serial ports. The microhard PX2 modules achieve self-organization, self-relay, and self-healing long-distance wireless communication between nodes through mesh networking.

[0019] The LPDDR4 storage module is an external X32 LPDDR4X 3733Mbps chip, and the MLU220 SoC chip and the LPDDR4X memory chip are directly connected in a point-to-point topology.

[0020] The node collaboration module includes an Image Handle submodule and a Web Controller submodule. The Image Handle submodule is used to complete multi-source data preprocessing and payload interface adaptation. The Web Controller submodule is used to parse the Note field in the network packet, start the inference pipeline, manage batch processing, and fill in the batch size when the transaction ends.

[0021] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An edge-distributed low-latency intelligent cluster system, characterized in that, It includes a master node, multiple slave nodes, and a wireless mesh communication unit; the master node and slave nodes establish a communication connection through the wireless mesh communication unit to form a star-shaped distributed cluster architecture; both the master node and slave nodes are configured with an edge intelligent computing platform; the edge intelligent computing platform includes an AI computing unit, a power module, an RJ45 network port, a UART serial port, a 32GeMMCFLASH, and an LPDDR4 storage module; the master node has a built-in TCPCluster module, an environment awareness module, a link control module, and a node collaboration module; the slave nodes have a built-in target detection module, a target recognition module, and a target decision module. The TCP Cluster module includes a TCP cluster manager thread and several TCP send threads, which are synchronized through a message queue. The TCP cluster manager thread listens for TCP connection requests from slave nodes, selects TCP send threads according to the load balancing strategy, and dynamically creates or destroys threads. The TCP send threads are responsible for bidirectional data transmission with the corresponding slave nodes. The load balancing strategy is based on dynamic task allocation according to the load status of slave nodes, and the specific processing steps include: S1B, Calculate the load factor of the slave node. The calculation formula is as follows: ,in, This refers to the load factor of the slave node; Let m be the length of the task queue of the m-th slave node; This represents the maximum length of the queue, with a value of 10. This represents the percentage of computing power used by the m-th slave node. This represents the total computing power of the slave nodes; m is the slave node number. S2B, task allocation strategy: The master node obtains information from each slave node in real time. ,choose The smallest slave node is assigned the computation task; if all slave nodes... When necessary, the model will automatically be downgraded to reduce the computational load of a single task and ensure low-latency operation of the cluster. The link control module includes a network bandwidth awareness submodule and a task data adaptive allocation submodule. The network bandwidth awareness submodule queries the current Mesh network's communication bandwidth and transmission rate in real time using AT commands from the microhard PX2 module. The task data adaptive allocation submodule dynamically matches a deep learning model based on the network bandwidth awareness results and the number of online slave nodes, and adaptively adjusts the resolution of the image data to be transmitted, thereby achieving dynamic matching of transmission link bandwidth pressure. The specific processing includes: S1C, Bandwidth Acquisition: Commands are sent via the microhard PX2 module of the wireless mesh communication unit to obtain the real-time network transmission rate. The average bandwidth is calculated using a moving average filtering algorithm. The calculation formula is as follows: ,in, Average bandwidth; is the real-time network transmission rate for the k-th data acquisition; N is the number of data acquisitions. This is the index for the number of data collections; k ranges from 0 to 4. S2C, bandwidth tiering and binding to corresponding image resolutions and learning models: when At Mbps, it corresponds to 800×600 resolution and VGG16-Small model; When 15 Mbps < At Mbps, this corresponds to a 600×512 resolution and the VGG16-Middle model; when At Mbps, it corresponds to a 224×224 resolution and the VGG16-Large model; S3C, Model Switching Decision: Calculated using the delayed prediction model formula, the formula is as follows: ,in, is the end-to-end delay prediction value for a single frame image; W is the image width; H is the image height; This refers to the total computing power of the MLU220 SoC chip. when When the resolution is reduced and the model is switched, the slave nodes are notified to preload the target model. The environmental perception module collects computing load, execution latency, and bandwidth fluctuations, and calculates the state change rate using the following formula: ,in, The rate of change of state; The real-time value of the parameter at the current time t: These are the historical parameter values ​​for the previous 10 monitoring periods; When the rate of change of any one of the parameters—computing load, execution latency, or bandwidth fluctuation—satisfies |ΔX|>0.1, feedback is sent to the node collaboration module and link control module to optimize task allocation and model switching. The collaborative working relationship among the modules is as follows: the wireless Mesh communication unit collects network bandwidth and task data and transmits it to the link control module and the environment perception module; the TCPCluster module manages node connections and thread scheduling, and the node collaboration module performs computation task allocation; the AI ​​computing unit completes parallel inference computation and returns the results; the environment perception module monitors node load and network status in real time, and optimizes task allocation and model switching strategies in reverse to achieve low-latency intelligent computing in unstable network environments.

2. The edge-distributed low-latency intelligent cluster system as described in claim 1, characterized in that: The AI ​​computing unit uses the MLU220 SoC chip, integrating a quad-core ARM Cortex-A55 processor and an NPU deep learning processor. It runs on the Ubuntu 16.04 operating system and the Neuware SDK software stack, and is used for lightweight deep learning model parallel inference. The processing flow includes: S1A, Model Loading and Initialization: Load the lightweight deep learning model and resolve the model parallelism through the CNRT runtime library. , The value can be 1 or 4; data parallelism ,satisfy Initialize three independent streams: data copying, model inference, and result return. S2A, Inference Calculation: The received image data is processed using formulas... Normalization, in which, The image data before normalization. The image data is normalized. Parallel inference is performed based on the MLU220SoC chip core allocation strategy. The formula for calculating the inference time is as follows: , , ,in, The reasoning process takes time; The time taken to push data onto the stack; To calculate the time consumed; This refers to the time spent popping data from the stack. The data volume is represented by FLOPs, and the floating-point operation volume is represented by FLOPs. The inference time is input as an execution delay parameter to the environment perception module for node status monitoring and closed-loop optimization, and at the same time provides a computational basis for the delay prediction model of the link control module. S3A. Post-processing of results: The non-maximum suppression algorithm is used to filter duplicate recognition results. The calculation formula is as follows: ,in, This is a non-maximum suppression algorithm; To retrieve the index corresponding to the maximum value; Let be the confidence level of the i-th detection box; For intersection, union, and comparison; For the i-th detection box; For the j-th detection box; For all j and j not equal to i; i and All of these are detection box numbers.

3. The edge-distributed low-latency intelligent cluster system as described in claim 2, characterized in that: The wireless Mesh communication unit uses a microhard PX2 module, which supports 802.11b / g / n high-speed wireless communication mode and four networking modes: AP / Client / Repeater / Mesh. It provides dual gigabit Ethernet ports, an RS232 serial port, and a USB 3.0 interface. The MLU220 SoC chip and the microhard PX2 module are interconnected through gigabit or 100 Mbps Ethernet ports and serial ports. The microhard PX2 modules achieve self-organization, self-relay, and self-healing long-distance wireless communication between nodes through Mesh networking.

4. The edge-distributed low-latency intelligent cluster system as described in claim 3, characterized in that: The LPDDR4 storage module is an external X32 LPDDR4X 3733Mbps chip, and the MLU220 SoC chip and the LPDDR4X memory chip are directly connected in a point-to-point topology.

5. The edge-distributed low-latency intelligent cluster system as described in claim 1, characterized in that, The node collaboration module includes an Image Handle submodule and a Web Controller submodule. The Image Handle submodule is used to complete multi-source data preprocessing and payload interface adaptation. The Web Controller submodule is used to parse the Note field in the network packet, start the inference pipeline, manage batch processing, and fill in the batch size when the transaction ends.

Images