Cloud edge end task processing method and system based on algorithm combination

Abstract

The invention provides a cloud edge-end task processing method and system based on algorithm combination, the system comprises a cloud side, an edge side and an end side for realizing a data processing flow, the cloud side is used for realizing configuration strategy input and decomposition and algorithm result display, the edge side is used for realizing an algorithm detection flow, and the end side utilizes an existing camera to realize algorithm detection. Performing algorithm enabling on a common monitoring camera which does not have the intelligent detection capability; the data processing flow comprises an issuing node A and an output node F which are located on the cloud side, and a frame extraction node B, a sensing node C, a processing node D and a logic node E which are located on the edge side, the four nodes on the edge side are four mutually independent modules, data circulation is achieved through algorithm scheduling, and the overall algorithm function is achieved in a module combination mode.

Description

Technical Field

[0001] The present invention relates to edge computing, and more particularly to a cloud-edge task processing method and system based on algorithm combination. Background Art

[0002] With the rapid development of the Internet, cloud computing has become an important way to manage data centers. However, with the rapid development of technologies such as the Internet of Things and artificial intelligence, data is generated and collected at an increasingly faster rate and in ever-larger quantities. In order to better process this data, it is necessary to perform computational processing where it is generated or close to where it is generated, and edge computing has emerged. Edge computing is a technology that moves computing and data storage to the edge of the network to improve response speed and reduce network bandwidth requirements.

[0003] Cloud-edge-end collaboration refers to the collaboration and cooperation between the three levels of cloud, edge, and end in a computing system. The cloud side refers to the cloud computing center, the edge refers to the edge computing node, and the end refers to the terminal device. In edge computing, cloud-edge-end collaboration can improve the processing speed and efficiency of the system, while also reducing network bandwidth requirements and latency.

[0004] The cloud computing center and edge computing nodes are closely connected. The cloud computing center can provide computing resources, data processing services, and storage services for edge computing nodes, and can also manage and control edge computing nodes. Edge computing nodes can perform computing and storage closer to terminal devices, thereby improving response speed and reducing network bandwidth requirements.

[0005] The relationship between edge computing nodes and terminal devices is closer. Edge computing nodes can perform calculations and storage near terminal devices, so that they can respond to requests from terminal devices more quickly. At the same time, they can also reduce the amount of data transmitted from terminal devices to the cloud side and reduce network bandwidth requirements.

[0006] For cloud-edge collaboration technology, how to design a task scheduling mechanism and a cloud-edge-end collaboration mechanism to maximize the utilization of computing resources and support more algorithm tasks and more analysis paths has become a key point. Summary of the invention

[0007] The present invention provides a cloud-edge-device task processing method and system based on algorithm combination, which solves the problem of maximizing the utilization of computing resources through a cloud-edge-device collaboration mechanism. The technical solution is as follows:

[0008] A cloud-edge-end task processing system based on algorithm combination, including a cloud side, an edge side and an end side for implementing a data processing flow, wherein the cloud side is used to implement configuration strategy input and decomposition, and algorithm result display, the edge side is used to implement an algorithm detection flow, and the end side uses existing cameras to perform algorithm empowerment on ordinary surveillance cameras that do not have intelligent detection capabilities; the data processing flow includes a sending node A and an output node F located on the cloud side, and a frame extraction node B, a perception node C, a processing node D and a logic node E located on the edge side, the four nodes on the edge side are four independent modules, the data flow is realized through algorithm scheduling, and the overall algorithm function is realized in the form of module combination.

[0009] When the user of the sending node A sends a task, it sends the overall function, including the specific sub-modules corresponding to the perception node C, the processing node D, and the logic node E.

[0010] The frame extraction node B is provided with a dynamic frame extraction function, and the frame extraction frequency is dynamically adjusted according to the priority of the task being sent, the number of tasks, and the processing time of each module.

[0011] The frame extraction frequency is adjusted as the tasks sent down by the cloud side are adjusted. Each adjustment uses the basic frame extraction frequency required by the highest priority task of each video as the frame extraction frequency of the video, that is:

[0012]

[0013] where f n It is expressed as the frame extraction frequency of the nth video channel. Indicates the basic frame extraction frequency required for the mth parsing task sent for the nth video.

[0014] The perception node C is used to realize the detection and attribute analysis functions of the target objects that the user is concerned about in the business scenarios required by the user. The cloud side updates and adds target detection and attribute recognition models based on the user's new business needs, and sends them to the edge at the same time. The perception node receives the model sent by the cloud side and the target attribute acquisition object corresponding to the perception node C decomposed by the sending node A, and performs corresponding target detection model activation and target detection.

[0015] The processing node D is used to implement the intermediate process processing in the overall function of the algorithm, including trajectory tracking, overlapping relationship between targets, and target pixel area calculation.

[0016] The logic node E implements various logic rule judgments based on the detection target output by the perception node C and the intermediate process state output by the processing node D.

[0017] The algorithm processing scheduling mechanism inside the edge side uses the logic of algorithm combination. The algorithm modules are independent of each other, and data flow is carried out in the form of parallel processing. The node that occupies the largest module of computing power resources and the longest processing time is used as the main node, and the remaining processing nodes make full use of the frame extraction interval to perform calculation processing on the remaining nodes. Task scheduling is divided into two cases:

[0018] Algorithm node module-level scheduling: Use the frame extraction interval to process the low-priority algorithm nodes of the previous task with low computing resource usage and short computing time.

[0019] Algorithm overall module-level scheduling: use the frame extraction interval to process all nodes of the previous low-priority algorithm.

[0020] The algorithm processing scheduling mechanism inside the edge side adopts the form of algorithm module combination, and distinguishes the task and module processing priorities at the same time. The processing priority of the same task is sorted according to the internal processing order, that is, B>C>D>E. Different tasks are processed according to the priority of the task defined by the user, and the frame extraction interval is fully utilized to perform low-priority tasks or algorithm node module calculation processing. The overall processing time of the nth task using this method is:

[0021] TIME n =max(timeB,timeC n , timeD n ,timeE n )

[0022] The computing resources for the nth task are:

[0023] S n =max(SB,SC n , S.D. n , S.E. n )

[0024] When the user only issues one parsing task, the edge box can support the following parsing video channels:

[0025]

[0026] Among them, timeB is the time required for each frame extraction behavior of the frame extraction node, timeC n is the actual processing time of the nth model of the perception node, timeD n timeE is the actual processing time of the nth processing child node of the processing node, n The actual processing time of the nth logical child node of the logical node;

[0027] SB is the computing power required by the nth processing subnode of the frame extraction node, SC nThe computing resources required by the nth processing node of the sensing node, SD n The computing resources required for the nth processing node of the processing node, SE n The computing resources required for the nth logical sub-node of the logical node;

[0028] S is all the computing resources of the edge box.

[0029] When the user issues multiple tasks, the edge box can support the following video parsing channels:

[0030]

[0031] At the same time, in order to avoid data backlog caused by long algorithm processing time due to high frame extraction frequency and meet the needs of real-time processing, the optional frame extraction frequency must meet the following requirements:

[0032] f n >TIME n .

[0033] A cloud-edge task processing method based on algorithm combination includes the following steps:

[0034] S1: The sending node A on the cloud side sends tasks to the edge side. The tasks are automatically decomposed into specific sub-modules corresponding to the sensing node C, processing node D, and logic node E on the edge side on the cloud side.

[0035] S2: The frame extraction node B located at the edge side is equipped with a dynamic frame extraction function. The frame extraction frequency is dynamically adjusted according to the priority of the task sent, the number of tasks, and the processing time of each module;

[0036] S3: The sensing node C receives the model sent by the cloud side and the target attribute acquisition object corresponding to the sensing node C decomposed by the sending node A, and performs corresponding target object detection model activation and target detection;

[0037] S4: Processing node D is used to implement the intermediate process processing in the overall function of the algorithm, including trajectory tracking, overlapping relationship between targets, and target pixel area calculation;

[0038] S5: Logic node E implements various logic rule judgments based on the detection target output by sensing node C and the intermediate process status output by processing node D;

[0039] S6: Output node F implements result output and presents the effect on the cloud side.

[0040] The cloud-edge-terminal task processing method and system based on algorithm combination designs a cloud-edge-terminal collaboration mechanism and an edge-side computing resource scheduling mechanism based on algorithm combination, realizing more algorithm functions, more analysis paths, and more computing resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a structural diagram of the cloud-edge-end task processing system based on algorithm combination;

[0042] Figure 2 is a schematic diagram of the processing interval of the frame extraction node for tasks a, b, and c;

[0043] Figure 3 It is a scheduling diagram of traditional algorithm processing flow tasks;

[0044] Figure 4 This is a task scheduling diagram for the master node and other nodes. Figure 1 ;

[0045] Figure 5 This is a task scheduling diagram for the master node and other nodes. Figure 2 ;

[0046] Figure 6 It is a flowchart of the cloud-edge-end task processing method based on algorithm combination;

[0047] Figure 7 yes Figure 1 A partial enlargement of Figure 1 ;

[0048] Figure 8 yes Figure 1 A partial enlargement of Figure 2 ;

[0049] Fig. 9 yes Figure 1 A partial enlargement of Figure 3 ;

[0050] Fig.10 yes Figure 1 A partial enlargement of Figure 4 . DETAILED DESCRIPTION

[0051] like Figure 1 As shown, combined Figure 7-10 The cloud-edge-end task processing system based on algorithm combination includes a cloud side, an edge side, and an end side for implementing the data processing process. The cloud side is used to implement configuration strategy input and decomposition, and algorithm result display. The edge side is used to implement the algorithm detection process. The end side uses existing cameras to perform algorithm empowerment on ordinary surveillance cameras that do not have intelligent detection capabilities.

[0052] The data processing flow is divided into 6 nodes, mainly the sending node A, the frame extraction node B, the perception node C, the processing node D, the logic node E, and the output node F. The sending node A and the output node F are located on the cloud side, the frame extraction node B, the perception node C, the processing node D, and the logic node E are located on the edge side. The four nodes on the edge side are four independent modules. The data flow is realized through algorithm scheduling, and the overall algorithm function is realized in the form of module combination. The functions of each node are as follows:

[0053] Step 1 Sending nodes: When the user sends a task to the sending node A, it sends the overall function, including the specific sub-modules corresponding to the perception node C, processing node D, and logic node E.

[0054] For example, when a user issues a functional requirement of "not wearing a helmet in key areas", the cloud side breaks it down into two types: perception node C for obtaining the two target attributes of "pedestrian target detection and attribute recognition" and "other object target detection and attribute recognition" + processing node D for the intermediate process processing of "target overlap relationship" + logic node E for the logical rule judgment of "area detection logic".

[0055] Step 2 Frame extraction node: Frame extraction node B is equipped with a dynamic frame extraction function. The frame extraction frequency is dynamically adjusted according to the priority of the task issued, the number of tasks, and the processing time of each module to fully utilize the computing resources and avoid problems such as packet loss caused by data backlog.

[0056] The frame extraction frequency is adjusted as the tasks sent by the cloud are adjusted. Each adjustment is based on the basic frame extraction frequency required by the highest priority task of each video channel as the frame extraction frequency of the video channel, that is:

[0057]

[0058] where f n It is expressed as the frame extraction frequency of the nth video channel. Indicates the basic frame extraction frequency required for the mth parsing task sent for the nth video.

[0059] The edge sets the image distribution rules. For each parsing task, the actual image processing interval is:

[0060]

[0061] like Figure 2 As shown in the figure, for example, when a certain video sends three detection tasks a, b, and c, the basic frame extraction frequencies required by the three algorithm tasks are 1000 (frame extraction every 1000ms), 2000 (frame extraction every 2000ms), and 125 (frame extraction every 125ms), respectively. Then the actual frame extraction frequency f of the video is nThe time required for each frame extraction is recorded as timeB, and the computing power required is SB.

[0062] Step 3 Perception Node: Perception node C is mainly used to realize the detection and attribute analysis functions of the target objects that users are concerned about in the business scenarios required by users. The cloud side can update and add target detection and attribute recognition models based on the user's new business needs, and send them to the edge at the same time. The perception node receives the model sent by the cloud side and the target attribute acquisition object corresponding to the perception node C decomposed by the sending node A, and performs corresponding target detection model activation and target detection.

[0063] The perception node target attribute acquisition module includes but is not limited to: pedestrian target detection and attribute recognition, motor vehicle target detection and attribute recognition, non-motor vehicle target detection and attribute recognition, face target detection and attribute recognition, other object target detection and attribute recognition and other sub-modules.

[0064] For example, when the user sends tasks including "pedestrian wandering" and "not wearing a helmet in key areas" on the cloud side, the cloud-side task decomposition module splits the tasks sent by the user into two perception nodes C: "pedestrian target detection and attribute recognition" and "other object target detection and attribute recognition (helmet)", which are recorded as C1 and C2 respectively. Each video channel is limited by the number of targets in the scene, and there are differences in the processing time of a single model. The basic processing time of the nth model of the perception node C is recorded as The actual processing time of the model is updated every m times, where m can be set by the user. Each update is calculated using the average processing time of the previous m models. The actual processing time of the nth model is timeC. n for:

[0065]

[0066] At the same time, the computing resources occupied by each sub-model are recorded as SC n ; i represents a natural number.

[0067] Step 4 Processing node: Processing node D is mainly used to implement the intermediate process processing in the overall function of the algorithm, mainly including trajectory tracking, overlapping relationship between targets, target pixel area calculation, etc. It plays a connecting role between perception nodes and logic nodes.

[0068] This node is limited by the number of targets output by the perception node in the previous step. There are differences in the processing time of a single processing sub-node. The basic processing time of the nth processing sub-node of the processing node D is recorded as The actual processing time of the processing sub-node is updated every m times, where m can be set by the user. Each update is calculated using the average processing time of the previous m times. The actual processing time of the nth processing sub-node is timeD n for:

[0069]

[0070] At the same time, the computing resources occupied by each sub-processing node are recorded as SD n ; i represents a natural number. The intermediate process of processing a node includes but is not limited to: target trajectory analysis, omitting the intermediate process (ie, no need to process the node), target overlap relationship, etc.

[0071] Step 5 Logic Node: Logic node E is mainly used to implement various logic judgment rules, such as wandering behavior logic, retrograde behavior logic, retention detection logic, counting logic, area judgment logic, etc. It implements various logic rule judgments based on the detection target output by the perception node C and the intermediate process state output by the processing node D.

[0072] This node is limited by the number of target intermediate states output by the previous processing node. There are differences in the processing time of a single logical sub-node. The basic processing time of the nth logical sub-node of the logical node E is recorded as The actual processing time of the logical sub-node is updated every m times, where m is user-configurable. Each update is calculated based on the average processing time of the previous m times. The actual processing time of the nth logical sub-node is timeE n for:

[0073]

[0074] At the same time, the computing resources occupied by each sub-processing node are recorded as SE n ; i represents a natural number. The logic judgment rules in the logic node include but are not limited to: wandering behavior logic, retrograde behavior logic, omission detection logic (i.e., no need to perform logic node judgment), counting logic, area detection logic, timing logic, etc.

[0075] Step 6 Output node: Output node F realizes result output and presents the effect on the cloud side.

[0076] There are four types of nodes inside the edge side, namely, frame extraction node B, perception node C, processing node D, and logic node E. The internal task scheduling mechanism of the nodes is different from the traditional algorithm processing scheduling mechanism.

[0077] The traditional algorithm processing scheduling mechanism is as follows: The traditional algorithm processing process is to complete target detection, target tracking, logic rule judgment and other processes within a single algorithm, that is, target perception, intermediate process processing, and logic rule judgment are all included in the same algorithm. The flow of data needs to be processed serially. The processing time is the sum of the processing time of each part, and the required computing power resources are also the sum of the computing power resources of each part. That is, the overall processing time of the nth task is:

[0078] TIME n0 =timeB+timeC0+timeD0+timeE0

[0079] Among them, the computing resources of the nth task are:

[0080] S n0 =SB+SC0+SD0+SE0

[0081] When the user only issues one parsing task, the edge box can support the following parsing video channels:

[0082]

[0083] Among them, S is all the computing resources of the edge box.

[0084] When the user issues multiple tasks, the edge box can support the following video parsing channels:

[0085]

[0086] i represents a natural number, and n represents n tasks.

[0087] At the same time, in order to avoid data backlog caused by long algorithm processing time due to high frame extraction frequency and meet the needs of real-time processing, the optional frame extraction frequency must meet the following requirements:

[0088] f n >TIME n0

[0089] The traditional algorithm processing flow task scheduling diagram is as follows Figure 3 shown.

[0090] In the present invention, the algorithm processing scheduling mechanism inside the edge side is as follows: using the logic of algorithm combination, each algorithm module is independent of each other, data flow is carried out in the form of parallel processing, the internal logic scheduling of the edge side is flexibly configured, and the node with the largest module occupying computing power resources and the longest processing time is used as the main node. The remaining processing nodes make full use of the frame extraction interval to perform calculation processing on the remaining nodes. Task scheduling is mainly divided into two cases:

[0091] Algorithm node module-level scheduling: The scheduling diagram is as follows Figure 4 As shown, the algorithm node with low priority and low computing resource usage and short time consumption of the previous task is processed by using the frame extraction interval;

[0092] The overall module-level scheduling of the algorithm: The scheduling diagram is as follows Figure 5 As shown, all nodes of the previous low priority algorithm are processed using the frame extraction interval;

[0093] The present invention adopts the form of algorithm module combination, and distinguishes the task and module processing priorities at the same time. The processing priority of the same task is sorted according to the internal processing order, that is, B>C>D>E. Different tasks are processed according to the priority of the task defined by the user, and the frame extraction interval is fully utilized to perform low-priority tasks or algorithm node module calculation processing. The overall processing time of the nth task using this method is:

[0094] TIME n =max(timeB,timeC n , timeD n timeE n )

[0095] Where timeB is the time required for each frame extraction behavior of the frame extraction node, timeC n is the actual processing time of the nth model of the perception node, timeD n timeE is the actual processing time of the nth processing child node of the processing node, n The actual processing time of the nth logical child node of the logical node.

[0096] The computing resources for the nth task are:

[0097] S n =max(SB,SC n , S.D. n , S.E. n )

[0098] Among them, SB is the computing power resources required by the nth processing subnode of the frame extraction node, SC n The computing resources required by the nth processing node of the sensing node, SD n The computing resources required for the nth processing node of the processing node, SE n The computing resources required for the nth logical sub-node of the logical node.

[0099] When the user only issues one parsing task, the edge box can support the following parsing video channels:

[0100]

[0101] Among them, S is all the computing resources of the edge box.

[0102] When the user issues multiple tasks, the edge box can support the following video parsing channels:

[0103]

[0104] At the same time, in order to avoid data backlog caused by long algorithm processing time due to high frame extraction frequency and meet the needs of real-time processing, the optional frame extraction frequency must meet the following requirements:

[0105] f n >TIME n

[0106] Furthermore, the present invention also has a method of dynamically adjusting the frame extraction strategy. When there are many targets in the video, the number of targets to be processed by the perception node C, the processing node D and the logic node E increases, which will increase the processing time and even cause data backlog, so that the system crashes. In order to avoid system failures caused by data backlogs, the frame extraction strategy is adjusted in the form of dynamic adjustment of the frame extraction frequency to achieve full utilization of computing resources, while avoiding data loss caused by data backlogs, and even crashes.

[0107] The specific method is to record the processing time of each node in the previous m times, combined with the conditions that the frame extraction frequency must meet:

[0108] f n >TIME n , the frame extraction frequency of each task is designed to be:

[0109] f n =TIME n +θ

[0110] Among them, θ is a positive constant, which can be designed based on experience or set according to user needs. n It changes dynamically, so the frame extraction frequency is dynamically adjusted every m times.

[0111] A cloud-edge task processing method based on algorithm combination includes the following steps:

[0112] S1: The sending node A on the cloud side sends tasks to the edge side. The tasks are automatically decomposed into specific sub-modules corresponding to the sensing node C, processing node D, and logic node E on the edge side on the cloud side.

[0113] S2: The frame extraction node B located at the edge side is equipped with a dynamic frame extraction function. The frame extraction frequency is dynamically adjusted according to the priority of the task sent, the number of tasks, and the processing time of each module;

[0114] S3: The sensing node C receives the model sent by the cloud side and the target attribute acquisition object corresponding to the sensing node C decomposed by the sending node A, and performs corresponding target object detection model activation and target detection;

[0115] S4: Processing node D is used to implement the intermediate process processing in the overall function of the algorithm, including trajectory tracking, overlapping relationship between targets, and target pixel area calculation;

[0116] S5: Logic node E implements various logic rule judgments based on the detection target output by sensing node C and the intermediate process status output by processing node D;

[0117] S6: Output node F implements result output and presents the effect on the cloud side.

[0118] Furthermore, in the edge side of the present invention, the logic of algorithm combination is used, the algorithm modules are independent of each other, the data flow is carried out in the form of parallel processing, the internal logic scheduling of the edge side is flexibly configured, and the node with the largest module occupying computing power resources and the longest processing time is used as the main node. The remaining processing nodes make full use of the frame extraction interval to perform calculation processing on the remaining nodes. Task scheduling is mainly divided into two cases:

[0119] Algorithm node module-level scheduling: use the frame extraction interval to process the algorithm nodes with low priority and low computing resource usage and short computing time of the previous task;

[0120] Algorithm overall module-level scheduling: use the frame extraction interval to process all nodes of the previous low-priority algorithm.

[0121] Furthermore, the frame extraction node B has a dynamic frame extraction adjustment strategy. When there are many targets in the video, the number of targets to be processed by the perception node C, the processing node D and the logic node E increases, which will increase the processing time and even cause data backlog, resulting in system crashes. In order to avoid system failures caused by data backlogs, the frame extraction strategy is adjusted in the form of dynamic adjustment of the frame extraction frequency to achieve full utilization of computing resources, while avoiding data loss caused by data backlogs, or even crashes.

[0122] The specific method is to record the processing time of each node in the previous m times, combined with the conditions that the frame extraction frequency must meet:

[0123] f n >TIME n , the frame extraction frequency of each task is designed to be:

[0124] f n =TIME n +θ

[0125] Among them, θ is a positive constant, which can be designed based on experience or set according to user needs. n It changes dynamically, so the frame extraction frequency is dynamically adjusted every m times.

[0126] The present invention has the following advantages:

[0127] 1. Cloud-edge collaboration to give full play to the advantages of computing resources

[0128] The traditional method of only using central deployment will result in a waste of computing resources for small parks or mining areas with fewer video points. Therefore, for scenarios such as small and medium-sized parks, the cloud-edge collaboration approach is adopted, and the use of edge computing boxes can give full play to the value of computing resources.

[0129] The algorithm parsing tasks sent by users on the cloud side are overall requirements. The cloud side supports the splitting of the overall algorithm tasks, automatically splitting them into three node modules: perception node, processing node, and logic node. When a task sent by the user is found to be not included in any of the perception node, processing node, and logic node after splitting, a work order will be automatically generated to remind the operation and maintenance personnel to update the corresponding nodes, and the updated nodes will be sent to the edge side.

[0130] 2. The overall framework of algorithm combination supports more task generation

[0131] The edge-side algorithm is divided into perception nodes, processing nodes, and logic nodes. Different node modules can be combined arbitrarily. Compared with the overall processing of traditional algorithms, on the one hand, it can save algorithm development time, avoid duplication of work, and focus on modules that need to be optimized and improved; on the other hand, the algorithm combination form supports the arbitrary combination of different modules, which changes the previous form of single algorithm development and can support the generation of more algorithm detection tasks. At the same time, it supports the use of the same perception node + multiple processing nodes + multiple logic nodes. When this form is used, the model expenditure of the target detection perception node is saved, thereby saving computing resources; in addition, the algorithm combination form can make full use of the computing resource difference and processing time difference between different nodes, make full use of the frame extraction time interval, and flexibly schedule computing resources so that the edge box can support more video analysis tasks.

[0132] Based on the overall framework of algorithm combination, the edge box can receive any algorithm requirements issued by cloud-side users, and update or add original node content to different node models after being disassembled by the cloud side.

[0133] 3. Dynamic adjustment of analysis frequency to avoid data backlog while ensuring maximum computing resource utilization

[0134] When there are many targets in the video, the number of targets to be processed by the perception node C, processing node D and logic node E will increase, which will increase the processing time and even cause data backlog, resulting in system crashes. In order to avoid system failures caused by data backlogs, the frame extraction strategy is adjusted in the form of dynamic adjustment of the frame extraction frequency to achieve full utilization of computing resources, while avoiding data loss caused by data backlogs, or even crashes.

[0135] 4. Flexible algorithm task scheduling management mechanism

[0136] Using the logic of algorithm combination, each algorithm module is independent of each other, and data flow is carried out in the form of parallel processing. The internal computing task priority and the remaining available computing power resources on the edge side are used for logical scheduling and flexible configuration. The node with the largest module occupying computing power resources and the longest processing time is used as the main node, and the remaining processing nodes make full use of the frame extraction interval to perform calculations and processing on the remaining nodes, so as to fully utilize the computing power resources.

[0137] 5. The number of video analysis channels can be dynamically adjusted

[0138] Combined with the number of algorithm parsing tasks issued by cloud users and the computing resources consumed by each algorithm module, and based on the task scheduling management mechanism on the edge side, the number of algorithm parsing video channels supported by the edge box can be flexibly adjusted according to user needs and the computing requirements of each task, so as to achieve a dynamic balance between the number of parsing algorithms and the number of parsing channels, and thus achieve full use of the computing resources on the edge side.

[0139] 6. Multi-edge device clusters can meet the needs of diverse algorithm scenarios

[0140] When multiple edge devices are clustered, users can issue different algorithm analysis tasks to different edge devices. Since different perception nodes, processing nodes, and logic nodes are integrated inside the edge box, it can support differentiated algorithm analysis to meet the overall needs of users and differentiated scenario analysis needs.

Claims

1. A cloud-edge-end task processing system based on algorithm combination, characterized by: It includes the cloud side, edge side and terminal side for implementing the data processing flow. The cloud side is used to implement the configuration strategy input and decomposition, and the algorithm result display. The edge side is used to implement the algorithm detection process. The terminal side uses the existing cameras to empower the algorithms of ordinary surveillance cameras that do not have intelligent detection capabilities. The data processing flow includes the sending node A and the output node F located on the cloud side, and the frame extraction node B, the perception node C, the processing node D and the logic node E located on the edge side. The four nodes on the edge side are four independent modules. The data flow is realized through algorithm scheduling, and the overall algorithm function is realized in the form of module combination.

2. The cloud-edge-end task processing system based on algorithm combination according to claim 1, characterized in that: When the user of the sending node A sends a task, it sends the overall function, including the specific sub-modules corresponding to the perception node C, the processing node D, and the logic node E.

3. The cloud-edge-end task processing system based on algorithm combination according to claim 1 is characterized in that: The frame extraction node B is provided with a dynamic frame extraction function, and the frame extraction frequency is dynamically adjusted according to the priority of the task being sent, the number of tasks, and the processing time of each module.

4. The cloud-edge-end task processing system based on algorithm combination according to claim 3 is characterized in that: The frame extraction frequency is adjusted as the tasks sent down by the cloud side are adjusted. Each adjustment uses the basic frame extraction frequency required by the highest priority task of each video as the frame extraction frequency of the video, that is: where f n It is expressed as the frame extraction frequency of the nth video channel. Indicates the basic frame rate required for the mth parsing task sent for the nth video.

5. The cloud-edge-end task processing system based on algorithm combination according to claim 1, characterized in that: The perception node C is used to realize the detection and attribute analysis functions of the target objects that the user is concerned about in the business scenarios required by the user. The cloud side updates and adds target detection and attribute recognition models based on the user's new business needs, and sends them to the edge at the same time. The perception node receives the model sent by the cloud side and the target attribute acquisition object corresponding to the perception node C decomposed by the sending node A, and performs corresponding target detection model activation and target detection.

6. The cloud-edge-end task processing system based on algorithm combination according to claim 1, characterized in that: The processing node D is used to implement the intermediate process processing in the overall function of the algorithm, including trajectory tracking, overlapping relationship between targets, and target pixel area calculation.

7. The cloud-edge-device task processing system based on algorithm combination according to claim 1, characterized in that: The logic node E implements various logic rule judgments based on the detection target output by the perception node C and the intermediate process state output by the processing node D.

8. The cloud-edge-end task processing system based on algorithm combination according to claim 1, characterized in that: The algorithm processing scheduling mechanism inside the edge side uses the logic of algorithm combination. The algorithm modules are independent of each other, and data flow is carried out in the form of parallel processing. The node that occupies the largest module of computing power resources and the longest processing time is used as the main node, and the remaining processing nodes make full use of the frame extraction interval to perform calculation processing on the remaining nodes. Task scheduling is divided into two cases: Algorithm node module-level scheduling: Use the frame extraction interval to process the low-priority algorithm nodes of the previous task with low computing resource usage and short computing time. Algorithm overall module-level scheduling: use the frame extraction interval to process all nodes of the previous low-priority algorithm.

9. The cloud-edge-end task processing system based on algorithm combination according to claim 8, characterized in that: The algorithm module combination is adopted, and the task and module processing priorities are distinguished at the same time. The processing priority of the same task is sorted according to the internal processing order, that is, B>C>D>E. Different tasks are processed according to the priority of the task defined by the user, and the frame extraction interval is fully utilized to perform low-priority tasks or algorithm node module calculations. Using this method, the overall processing time of the nth task is: TIME n smax(timeB,timeC n ,timeD n ,timeE n ) The computing resources for the nth task are: S n =max(SB,SC rv ,SD rv ,SE n ) When the user only issues one parsing task, the edge box can support the following parsing video channels: Among them, timeB is the time required for each frame extraction behavior of the frame extraction node, and timeC n is the actual processing time of the nth model of the perception node, timeD n timeE is the actual processing time of the nth processing child node of the processing node, n The actual processing time of the nth logical child node of the logical node; SB is the computing power required by the nth processing subnode of the frame extraction node, SC n The computing resources required by the nth processing node of the sensing node, SD n The computing resources required for the nth processing node of the processing node, SE n The computing resources required for the nth logical sub-node of the logical node; S is all computing resources of the edge box; When the user issues multiple tasks, the edge box can support the following video parsing channels: At the same time, in order to avoid data backlog caused by long algorithm processing time due to high frame extraction frequency and meet the needs of real-time processing, the optional frame extraction frequency must meet the following requirements: f n >TIME n 。 10. A cloud-edge task processing method based on algorithm combination, comprising the following steps: S1: The sending node A on the cloud side sends tasks to the edge side. The tasks are automatically decomposed into specific sub-modules corresponding to the sensing node C, processing node D, and logic node E on the edge side on the cloud side. S2: The frame extraction node B located at the edge side is equipped with a dynamic frame extraction function. The frame extraction frequency is dynamically adjusted according to the priority of the task sent, the number of tasks, and the processing time of each module; S3: The sensing node C receives the model sent by the cloud side and the target attribute acquisition object corresponding to the sensing node C decomposed by the sending node A, and performs corresponding target object detection model activation and target detection; S4: Processing node D is used to implement the intermediate process processing in the overall function of the algorithm, including trajectory tracking, overlapping relationship between targets, and target pixel area calculation; S5: Logic node E implements various logic rule judgments based on the detection target output by sensing node C and the intermediate process status output by processing node D; S6: Output node F implements result output and presents the effect on the cloud side.

Images