A multi-target detection method based on size model cooperation and distributed edge computing

Abstract

This invention discloses a multi-object detection method based on large-scale model collaboration and distributed edge computing, belonging to the fields of edge computing and computer vision. The input image is segmented into multiple sub-patterns, establishing a mapping relationship between local and global coordinates. Multiple edge nodes use small object detection models for inference, outputting target bounding boxes, category confidence scores, and feature vectors. The master node maps the detection results to global coordinates, constructs a cross-pattern candidate matching set, and determines and fuses these based on spatial constraints and feature similarity to achieve target identity association, re-identification, and truncated target reconstruction. For targets with low confidence, patch boundaries, and ambiguous matching, the master node calls the large object detection model for secondary detection and correction as needed, outputting globally consistent detection results and target identity identifiers. This method reduces the impact of duplicate targets, target truncation, and identity association errors on the results, lowers the master node's computational load and data transmission volume, and balances real-time performance and accuracy.

Description

Technical Field

[0001] This invention belongs to the field of edge computing and computer vision technology, and specifically relates to a multi-target detection method based on size model collaboration and distributed edge computing. Background Technology

[0002] In edge computing scenarios, multi-object detection tasks exhibit characteristics such as variable perspectives, easily truncated targets, and limited computing power of single nodes. In the edge-device paradigm, object detection typically faces complex heterogeneous environments, such as significant differences in computing resources among edge devices, network bandwidth bottlenecks, and drastic changes in target morphology across multiple perspectives. These complex scenario requirements necessitate that multi-object detection not only possess high real-time response capabilities but also achieve high-precision target recognition and continuous tracking.

[0003] However, current solutions struggle to balance low latency and high accuracy, and single-point device detection often performs poorly for truncated targets that cross field of view or are located at image edges. Therefore, there is an urgent need to design an efficient distributed multi-target detection method that not only meets the real-time requirements at the edge but also achieves high-precision target detection and cross-regional re-identification in complex scenarios through multi-node collaboration and intelligent correction.

[0004] Collaborative reasoning between small and large models and edge detection is a crucial research direction for large-scale video surveillance and dynamic tracking. Traditional centralized object detection methods typically transmit all high-definition video streams or image data collected at the edge to the cloud or central nodes for centralized processing. This method consumes a significant amount of network bandwidth and is prone to high transmission latency, making it difficult to apply effectively in collaborative tasks with extremely high real-time requirements, such as dynamic tracking. To address the challenges of latency and bandwidth, existing edge computing solutions often deploy small object detection models independently on each edge node. However, existing technologies suffer from several drawbacks. First, limited by the computing resources and memory bottlenecks of edge devices, lightweight models have weak feature extraction capabilities, making them prone to missed detections or low-confidence misjudgments when dealing with complex backgrounds, small-scale targets, or targets truncated at the edge of the field of view. Second, each edge node executes detection reasoning independently, lacking a global collaborative perception mechanism for targets across fields of view and across map tiles. When the same target moves between the fields of view of different nodes or is segmented, the lack of unified scheduling of spatial location and features often leads to duplicate target identification or loss of identity, resulting in inconsistent global detection results. Summary of the Invention

[0005] To address the limitations of lightweight models in edge node detection accuracy, inconsistent target identities across graph tiles, and low global resource scheduling efficiency, this invention aims to provide a multi-target detection method based on large and small model collaboration and distributed edge computing. It seeks to achieve efficient and globally consistent multi-target detection through a collaborative mechanism between lightweight edge detection and large model correction at the master control end.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a multi-target detection method based on size model collaboration and distributed edge computing, comprising the following steps:

[0007] Step 1: Construct an image spatial mapping and slicing model, divide the input image into multiple sub-tiles according to a preset partitioning strategy, and establish the mapping relationship between the sub-tiles and global coordinates;

[0008] Step 2: Deploy small object detection models on multiple edge computing nodes, perform object detection on the sub-map tiles respectively, and output the bounding box, class confidence and object feature vector of each object;

[0009] Step 3: Upload the detection results of each edge node to the master node, perform unified mapping of the detection results based on global coordinates, and construct a cross-tile target candidate matching set;

[0010] Step 4: Design a cross-tile target fusion mechanism based on spatial constraints and feature similarity to determine the consistency of candidate matching targets and realize cross-regional target re-identification and target reconstruction;

[0011] Step 5: Deploy a large target detection model on the master node, perform secondary correction on low-confidence targets and targets in the boundary area, and obtain globally consistent multi-target detection results;

[0012] Step 6: Output the target detection results and target identification in a unified coordinate system.

[0013] Optionally, the execution process of step 1 includes the following steps:

[0014] Step 1.1: Perform coverage segmentation of the spatial region based on the preset segmentation strategy. Let the input image be... Divide it into The input image is equivalent to the union of all sub-plots, as shown in the formula:

[0015]

[0016] in, Indicates the first There are 10 sub-tiles; and adjacent sub-tiles satisfy the following: ;

[0017] Step 1.2: Establish the global coordinate mapping relationship for each sub-tile, using the following formula:

[0018]

[0019]

[0020] in, For local coordinates of the subgraph, Global coordinates For the first Spatial offset of each sub-plot;

[0021] Optionally, the execution process of step 2 includes the following steps:

[0022] Step 2.1: Output the detection result through the target detection mini-model, expressed as:

[0023]

[0024] in, For bounding box, For category labels, For confidence level, For the target feature vector, This represents the number of targets in the sub-tile.

[0025] Optionally, the execution process of step 4 includes the following steps:

[0026] Step 4.1: Define the target matching scoring function, the formula is:

[0027]

[0028] in, With the target center point, , , These are the weighting coefficients. Spatial scale parameter;

[0029] Step 4.2: Set matching criteria; when the conditions are met... At that time, determine the target With the goal They belong to the same entity.

[0030] Step 4.3: Construct an undirected graph , where the set of nodes edge set The target clustering results were obtained through connected component analysis. Each of them This represents a set of detections for the same target.

[0031] Step 4.4: Cross-tile target fusion. Define the joint decision function as follows:

[0032]

[0033] When the judgment result satisfies At that time, the cross-tile target fusion mechanism is executed.

[0034] Optionally, the execution process of step 5 includes the following steps:

[0035] Step 5.1: Obtain the detection confidence level And set a judgment threshold. ;

[0036] Step 5.2: Determine the target state; when the target state is satisfied... Or the target is located in the boundary area of ​​the cut block. At that time, the main control terminal's target detection large model is triggered to perform secondary detection;

[0037] Step 5.3: For targets that meet the correction conditions, the main control node calls the target detection large model to perform high-precision secondary detection, improving detection accuracy and consistency. The formula is:

[0038]

[0039] in, This represents a large-scale object detection model. This represents the image region centered on the target.

[0040] This invention provides a multi-target detection method based on small-scale model collaboration and distributed edge computing. Addressing the challenges of limited computational resources and inconsistent target identities across different map tiles in multi-target detection tasks within edge computing scenarios, this invention proposes a collaborative detection method that integrates small-scale edge model detection with high-precision master-control correction.

[0041] In distributed multi-view detection or large-scale image block processing, existing solutions often employ a processing mode where each computing node infers independently, and a central node centrally aggregates the results. The detection boxes output by each node are typically established in different local coordinate systems. Without coordinate transformation, target association, and result verification processes, the same target may be repeatedly identified in adjacent blocks, or information may be missing due to its location at the block edge. While centralized detection facilitates unified processing, it increases image transmission volume and the computational burden on the central node; conversely, completely independent edge detection makes it difficult to guarantee the consistency of local results across the entire image. Therefore, existing methods still exhibit a significant constraint between detection latency, computational overhead, and result accuracy.

[0042] To address the aforementioned issues, this invention records the spatial position of each sub-plot in the original image during the image segmentation stage, and accordingly restores the local detection boxes output by the edge nodes to a unified global coordinate system. For detection results in adjacent plots that may belong to the same target, the spatial overlap, proximity, and feature vector similarity between targets are considered for association judgment. A graph structure is used to describe the matching relationship between candidate targets to complete the merging of duplicate detection results, cross-plot identity association, and truncated target reconstruction. For targets with confidence levels below a set threshold, close to plot boundaries, or with ambiguous associations, the corresponding image regions are submitted to the master node for supplementary inference and result correction by the target detection large model, thereby reducing the computational and transmission overhead caused by the large model repeatedly processing all image data.

[0043] This invention significantly improves the overall accuracy and consistency of multi-target detection in complex scenes, effectively solves the problems of target identity loss and duplicate detection across map tiles, and can adapt to the high-performance detection requirements in large-scale distributed visual perception scenarios. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the multi-target detection method based on large and small model collaboration and distributed edge computing of the present invention.

[0045] Figure 2 This is a flowchart of the multi-target detection method based on large and small model collaboration and distributed edge computing of the present invention. Detailed Implementation

[0046] The present invention will be described in detail below with reference to specific implementation examples. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention.

[0047] like Figure 1 As shown, this invention provides a multi-target detection method based on size model collaboration and distributed edge computing. The method steps are as follows:

[0048] S1 constructs an image spatial mapping and slicing model, which divides the input image into multiple sub-blocks according to a preset partitioning strategy, and establishes a mapping relationship between the sub-blocks and global coordinates;

[0049] S2: Deploy small object detection models on multiple edge computing nodes, perform object detection on the sub-map tiles respectively, and output the bounding box, class confidence and object feature vector of each object;

[0050] S3: Upload the detection results of each edge node to the master node, perform unified mapping of the detection results based on global coordinates, and construct a cross-tile target candidate matching set;

[0051] S4: Design a cross-tile target fusion mechanism based on spatial constraints and feature similarity to determine the consistency of candidate matching targets and realize cross-regional target re-identification and target reconstruction;

[0052] S5: Deploy a large target detection model on the master node to perform secondary correction on low-confidence targets and targets in the boundary area to obtain globally consistent multi-target detection results;

[0053] S6: Output the target detection results and target identification in a unified coordinate system.

[0054] The specific execution steps and procedures are as follows: Figure 2 As shown, the following description, in conjunction with specific embodiments and execution steps, will provide further details:

[0055] The execution process of step S1 can be broken down into the following steps:

[0056] (1.1) Image Overlapping Segmentation. Based on the preset segmentation strategy, the spatial region is segmented to cover the image. Let the input image be... Divide it into The input image is equivalent to the union of all sub-plots, as shown in the formula:

[0057]

[0058] in, Indicates the first There are [number] sub-tiles. Adjacent sub-tiles also satisfy: ;

[0059] (1.2) Establish global coordinate mapping. Model the global coordinate mapping relationship corresponding to each sub-tile, with the following formula:

[0060]

[0061]

[0062] in, For local coordinates of the subgraph, Global coordinates For the first Spatial offset of each sub-plot;

[0063] The execution process of step S2 can be broken down into the following steps:

[0064] (2.1) Small-model object detection. Each edge node receives a sub-map patch, performs inference through a small-model object detection, and outputs the detection result, represented as:

[0065]

[0066] in, For bounding box, For category labels, For confidence level, For the target feature vector, This represents the number of targets in the sub-tile.

[0067] Step S3: Upload the detection results of each edge node to the master node, perform unified mapping of the detection results based on global coordinates, and construct a cross-tile target candidate matching set.

[0068] The execution process of step S3 can be broken down into the following steps:

[0069] (3.1) Summary of detection results. Each edge computing node uploads the detection results of its sub-plots to the master node, forming a global set of detected targets D;

[0070] (3.2) Global coordinate unification. Based on the coordinate mapping relationship established in step S1, the detection results in each sub-plot are mapped from the local coordinate system to a unified global coordinate system to obtain a unified target set;

[0071] (3.3) Candidate matching set for cross-tile targets. For targets located in different sub-tiles but spatially adjacent or overlapping areas, a cross-tile target candidate matching set P is constructed based on their spatial proximity and feature similarity, which is used for subsequent target consistency determination and fusion processing.

[0072] The execution process of step S4 includes the following operations:

[0073] (4.1) Matching score calculation. For target pairs in the candidate matching set P, a target matching score function is defined, with the following formula:

[0074]

[0075] in, With the target center point, , , These are the weighting coefficients. Spatial scale parameter;

[0076] (4.2) Consistency Determination. Set matching criteria; when these criteria are met... At that time, determine the target With the goal They belong to the same entity.

[0077] (4.3) Target Clustering. Construct an undirected graph. , where the set of nodes edge set The target clustering results were obtained through connected component analysis. Each of them This represents a set of detections for the same target.

[0078] (4.4) Cross-tile target fusion. Define the joint decision function, the formula is:

[0079]

[0080] When the judgment result satisfies At that time, the cross-tile target fusion mechanism is executed.

[0081] The execution process of step S5 includes the following operations:

[0082] (5.1) Target confidence acquisition. Extract the detection confidence of each target in the unified target set. And set a judgment threshold. ;

[0083] (5.2) Target Status Determination. Based on the target confidence level and its spatial location, targets are screened. When the target status is satisfied... Or the target is located in the boundary area of ​​the cut block. At that time, the main control terminal's target detection large model is triggered to perform secondary detection;

[0084] (5.3) Secondary detection trigger: For targets that meet the correction conditions, the main control node calls the target detection large model to perform high-precision secondary detection, improving detection accuracy and consistency. The formula is: in, This represents a large-scale object detection model. This represents the image region centered on the target.

[0085] Compared with the prior art, the present invention has at least the following beneficial effects:

[0086] First, this invention effectively balances detection accuracy and inference latency. By deploying a lightweight model at the edge to handle most common targets, it only entrusts low-confidence and truncated targets to the large model at the main control end for secondary correction, significantly reducing the computational load and communication bandwidth requirements of the main control node. Second, it achieves excellent global consistency perception by introducing a cross-plot target fusion mechanism based on spatial constraints and feature similarity, solving the target splitting and ID duplication problems caused by traditional block detection. Third, this invention has strong adaptability to heterogeneous edge environments and can be flexibly deployed in various distributed edge computing nodes, providing solid technical support for real-time multi-target monitoring and tracking in large-scale scenarios.

[0087] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-target detection method based on size model collaboration and distributed edge computing, characterized in that, Includes the following steps: Step 1: Construct an image spatial mapping and slicing model, divide the input image into multiple sub-tiles according to a preset partitioning strategy, and establish the mapping relationship between the sub-tiles and global coordinates; Step 2: Deploy small object detection models on multiple edge computing nodes, perform object detection on the sub-map tiles respectively, and output the bounding box, class confidence and object feature vector of each object; Step 3: Upload the detection results of each edge node to the main control node, perform unified mapping of the detection results based on global coordinates, and construct a cross-tile target candidate matching set; Step 4: Design a cross-tile target fusion mechanism based on spatial constraints and feature similarity, perform consistency judgment on candidate matching targets, and realize cross-regional target re-identification and target reconstruction; Step 5: Deploy a large target detection model on the master node, perform secondary correction on low-confidence targets and targets in the boundary area, and obtain globally consistent multi-target detection results; Step 6: Output the target detection results and target identification in a unified coordinate system.

2. The multi-target detection method based on size model collaboration and distributed edge computing as described in claim 1, characterized in that, The execution process of step 1 includes the following operations: Step 1.1: Perform coverage segmentation of the spatial region based on the preset segmentation strategy. Let the input image be... Divide it into The input image is equivalent to the union of all sub-plots, as shown in the formula: in, Indicates the first There are 10 sub-tiles; and adjacent sub-tiles satisfy the following: ; Step 1.2: Establish the global coordinate mapping relationship for each sub-tile, using the following formula: in, For local coordinates of the subgraph, Global coordinates For the first Spatial offset of each sub-tile.

3. The multi-target detection method based on size model collaboration and distributed edge computing as described in claim 1, characterized in that, The execution process of step 2 includes the following operations: Step 2.1: Output the detection result through the target detection mini-model, as follows: in, For bounding box, For category labels, For confidence level, For the target feature vector, This represents the number of targets in the sub-tile.

4. The multi-target detection method based on size model collaboration and distributed edge computing as described in claim 1, characterized in that, The execution process of step 4 includes the following operations: Step 4.1: Define the target matching scoring function, the formula is: in, With the target center point, , , These are the weighting coefficients. Spatial scale parameter; Step 4.2: Set matching criteria; when the conditions are met... At that time, determine the target With the goal Belonging to the same entity; Step 4.3: Construct an undirected graph , where the set of nodes edge set The target clustering results were obtained through connected component analysis. Each of them A set of detections representing the same target; Step 4.4: Define the joint decision function, the formula is: When the judgment result satisfies At that time, the cross-tile target fusion mechanism is executed.

5. The multi-target detection method based on size model collaboration and distributed edge computing as described in claim 1, characterized in that, The execution process of step 5 includes the following operations: Step 5.1: Obtain the detection confidence level And set a judgment threshold. ; Step 5.2: Determine the target state; when the target state is satisfied... Or the target is located in the boundary area of ​​the cut block. At that time, the main control terminal's target detection large model is triggered to perform secondary detection; Step 5.3: For targets that meet the correction conditions, the main control node calls the target detection large model to perform high-precision secondary detection, improving detection accuracy and consistency. The formula is: in, This represents a large-scale object detection model. This represents the image region centered on the target.

Images