A target detection method and system adaptive to evolution of an inference platform

Abstract

The application discloses a target detection method and system adaptive to evolution of an inference platform, and the target detection method comprises the following steps: acquiring a to-be-detected image on a board-mounted device, inputting the to-be-detected image into a target detection model for inference, performing target preliminary screening and multi-dimensional feature output; when the target feature changes, performing normalization processing on a feature embedding vector of a candidate target, and performing similarity matching with a category prototype; enhancing target detection confidence based on the prototype matching similarity, and generating enhanced confidence; performing target screening and non-maximum suppression based on the enhanced confidence, and obtaining a detection result; and updating the corresponding category prototype according to a target sample meeting a condition in the detection result. By introducing a category prototype representation, similarity matching is performed between the feature embedding of the candidate target and the category prototype, the detection confidence is re-evaluated by using the prototype matching result, and the detection capability for the feature change target is improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and more particularly to a target detection method and system that adaptively evolves on an inference platform. Background Technology

[0002] Object detection is a crucial technology in computer vision, widely applied in scenarios such as remote sensing monitoring, industrial defect detection, security monitoring, and intelligent equipment perception. As artificial intelligence technology advances towards real-time and on-site implementation, object detection algorithms are increasingly deployed on edge devices, such as embedded computing platforms, industrial terminals, unmanned systems, and low-power intelligent devices.

[0003] Edge object detection is typically limited by computing power, storage space, and energy consumption. After deployment, models often focus on inference, making frequent offline retraining or cloud updates difficult. However, the actual operating environment at the edge is highly uncertain. Imaging conditions, background distribution, target appearance, and noise characteristics may change continuously over time, leading to a gradual decline in detection performance based on fixed-parameter models.

[0004] Therefore, achieving adaptive evolution of target detection models on edge inference platforms has become a crucial technical problem that urgently needs to be solved in the field of edge intelligence.

[0005] Existing edge object detection solutions typically employ lightweight convolutional neural networks or compressed detection models, which are trained offline and then deployed to edge devices for forward inference. During the inference phase, these solutions only output the bounding box location, object confidence score, and category information, while the model parameters remain static.

[0006] Some technologies attempt to improve performance through periodic model replacement, remote cloud updates, or incremental training, but these often rely on manually labeled data, stable network connections, or additional computing power, making them difficult to apply directly in edge scenarios where computing power is limited and real-time requirements are high.

[0007] The disadvantages of existing technologies include: Target detection models deployed on edge-end onboard devices can typically perform target detection tasks correctly when the target appearance remains consistent with that during training. However, when the characteristics of the target change during actual operation, such as due to damage, occlusion, deformation, pose changes, or local missing parts, the target appearance differs significantly from the original training samples, and the detection confidence and accuracy of existing detection models often decrease significantly.

[0008] In existing technologies, addressing the problems usually requires re-collecting samples and performing offline retraining or model updates. This approach not only relies on additional computing resources but is also difficult to implement in real time in onboard inference environments, failing to meet the needs of edge devices to adapt to changes in the characteristics of the target being tested. Summary of the Invention

[0009] In view of the above problems, the present invention is proposed to provide an adaptively evolving target detection method and system on an inference platform that overcomes or at least partially solves the above problems.

[0010] According to one aspect of the present invention, an adaptively evolving target detection method on an inference platform is provided, the target detection method comprising: The onboard device acquires the image to be detected and inputs it into the target detection model for inference, performing initial target screening and multi-dimensional feature output; When the target features change, the feature embedding vector of the candidate target is normalized and similarity matching is performed with the category prototype. The confidence score of target detection is enhanced based on prototype matching similarity, and an enhanced confidence score is generated. Target screening and nonmaximum suppression are performed based on the enhanced confidence level to obtain the detection results; Based on the target samples that meet the conditions in the detection results, the corresponding category prototypes are updated.

[0011] Optionally, the step of acquiring the image to be detected on the onboard device and inputting it into the target detection model for inference, performing initial target screening and multi-dimensional feature output specifically includes: The image to be detected is acquired on the onboard device and input into the target detection model for inference; Output candidate bounding boxes, target confidence scores, and corresponding feature embedding vectors; Determine whether the target detection confidence level is lower than a preset threshold, which is used to characterize whether the target features have changed.

[0012] Optionally, the initial screening of targets and the output of multidimensional features specifically include: Image preprocessing: The input image is scaled proportionally and its boundaries are padded, and then normalized and converted to match the dimensions of the model input. Multi-scale feature extraction: Forward inference is performed using a deep convolutional network to obtain raw output tensors with multiple sizes; Dimension Reshaping and Parameter Parsing: Reshape the original tensor into (B, na, ny, nx, no) format to restore the anchor box dimensions; Candidate box decoding: Apply the Sigmoid activation function to the output parameters; Perform coordinate transformation using preset grid offset and anchor frame parameters to calculate the center coordinates (xy) and width / height (wh) of the target candidate box: Feature separation and extraction: Simultaneously segment the decoded tensor to obtain the target existence confidence, class probability distribution, and fixed-length feature embedding vector.

[0013] Optionally, the step of normalizing the feature embedding vector of the candidate target and performing similarity matching with the category prototype when the target features change specifically includes: Feature vector normalization: Perform L2 norm normalization on the feature embedding vector and the preset class prototype vector corresponding to each candidate box, and project the vectors onto the unit hypersphere. Similarity matrix operation: By calculating the product of the normalized feature vector and the prototype matrix, a cosine similarity matrix describing the degree of similarity between the target and each category is obtained; Category attribute prediction: The maximum value of each row in the similarity matrix is ​​retrieved as the maximum similarity score, and the corresponding category index is mapped to the pre-detection category label of the target.

[0014] Optionally, the enhancement of the target detection confidence based on prototype matching similarity, specifically including the generation of enhanced confidence, includes: Match validity determination: Perform a quantitative comparison between the maximum similarity score and the preset similarity threshold, and filter out low-relevance feature interference through a step function or threshold mask; Confidence fusion calculation: The original target existence confidence score is multiplied by the matching validity determination result. Confidence reconstruction: The initial detection confidence is enhanced or corrected by similarity constraints to suppress spurious targets with mismatched class features.

[0015] Optionally, the step of performing target screening and non-maximum suppression based on the enhanced confidence level to obtain the detection results specifically includes: Initial screening based on confidence level: Based on the enhanced confidence level, candidate targets with confidence levels below a preset threshold are eliminated; Non-maximum suppression: Residual candidate boxes are sorted in descending order of confidence, and redundant boxes with an overlap of higher than the suppression threshold are removed by iteratively calculating the overlap area. Geometric space mapping: The coordinates of the filtered candidate boxes are inversely scaled from the model inference space according to the preprocessing ratio, and boundary constraint clipping is performed to restore them to the original image physical coordinate system; Result encapsulation output: The output includes the final detection message containing the target category index, fusion confidence score, and original image coordinate information.

[0016] Optionally, updating the corresponding category prototype based on target samples that meet the conditions in the detection results specifically includes: High-confidence feature sampling: Selecting target feature vectors that meet the high-confidence threshold and have correct spatial matching from the detection results; Feature centroid calculation: Calculate the mean of high-confidence feature vectors belonging to the same category within the current batch to obtain the instantaneous feature representation of the corresponding category; Exponential Moving Average Update: Performs an exponential moving average update on the preset category prototypes using the momentum factor. Statistical counting synchronization: Accumulates the sample observation counts for the corresponding category, enabling robust tracking of target feature evolution under specific environments.

[0017] This invention also provides an adaptively evolving target detection system on an inference platform, applying the aforementioned adaptively evolving target detection method on an inference platform. The target detection system includes: The onboard inference and detection module is used to acquire the image to be detected on the onboard device, input it into the target detection model for inference, and perform initial screening of targets and output multi-dimensional features. The prototype matching module is used to normalize the feature embedding vector of the candidate target when the target features change, and to perform similarity matching with the category prototype. The confidence enhancement module is used to enhance the confidence of target detection based on prototype matching similarity and generate the enhanced confidence score. The results filtering module is used to perform target filtering and non-maximum suppression based on the enhanced confidence level to obtain the detection results. The prototype update module is used to update the corresponding category prototype based on the target samples that meet the conditions in the detection results.

[0018] This invention provides a target detection method and system that adaptively evolves on an inference platform. The target detection method includes: acquiring an image to be detected on an onboard device and inputting it into a target detection model for inference, performing initial target screening and multi-dimensional feature output; when target features change, normalizing the feature embedding vectors of candidate targets and performing similarity matching with category prototypes; enhancing the target detection confidence based on prototype matching similarity to generate enhanced confidence; performing target screening and non-maximum suppression based on the enhanced confidence to obtain detection results; and updating the corresponding category prototypes based on target samples that meet the conditions in the detection results. By introducing category prototype representation, matching the feature embeddings of candidate targets with category prototypes for similarity, and re-evaluating the detection confidence using the prototype matching results, the detection capability for targets with changing features is improved.

[0019] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart of an adaptive evolution target detection method on an inference platform provided by an embodiment of the present invention; Figure 2 A detailed flowchart of an adaptive evolution target detection method on an inference platform provided in an embodiment of the present invention. Detailed Implementation

[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0023] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0025] The purpose of this invention is to provide a target detection method suitable for onboard inference platforms, enabling the detection model to complete the target detection task normally on the board, and after the characteristics of the target being tested change (such as damage or occlusion), the target detection performance is dynamically improved during the inference stage by introducing a prototype matching mechanism, without the need to retrain the detection model.

[0026] This invention provides an adaptive evolution target detection method on an onboard inference platform, comprising: when the target has a normal appearance, performing detection using a conventional target detection process; when the characteristics of the target change due to damage, occlusion, or other factors, compensating and enhancing the detection results through a prototype matching mechanism to improve overall detection performance.

[0027] The object detection model running on the onboard device outputs candidate bounding boxes, object confidence scores, and corresponding feature embedding vectors during the inference phase. When the object appearance is consistent with that during training, the detection model can successfully complete object detection based on the object confidence scores.

[0028] When the target under test undergoes feature changes such as damage, occlusion, or partial loss, the target confidence output by the detection model may decrease. This invention introduces a category prototype representation, performs similarity matching between the feature embedding of candidate targets and category prototypes, and uses the prototype matching results to re-evaluate the detection confidence, thereby improving the detection capability for targets with changed features.

[0029] like Figure 1 As shown, an adaptively evolving target detection method on an inference platform is described, the target detection method comprising: The onboard device acquires the image to be detected and inputs it into the target detection model for inference, performing initial target screening and multi-dimensional feature output; When the target features change, the feature embedding vector of the candidate target is normalized and similarity matching is performed with the category prototype. The confidence score of target detection is enhanced based on prototype matching similarity, and an enhanced confidence score is generated. Target screening and nonmaximum suppression are performed based on the enhanced confidence level to obtain the detection results; Based on the target samples that meet the conditions in the detection results, the corresponding category prototypes are updated.

[0030] like Figure 2 As shown, an adaptively evolving target detection method on an inference platform includes: Acquire input images from onboard devices; Object detection model inference; Output candidate bounding boxes, detection confidence scores, and feature embeddings; If the confidence level is lower than the threshold, proceed with the following steps; otherwise, use the standard detection results and output the target detection results. Determine if the target's characteristics have changed (damaged / occluded / missing); Feature embedding normalization processing; Similarity matching is performed based on category prototypes; Detection confidence is enhanced based on similarity; Target screening and NMS based on enhanced confidence; Output the enhanced target detection results; Select high-confidence target samples; Update the corresponding category model.

[0031] The method includes at least the following steps: The image to be detected is acquired on the onboard device and input into the target detection model for inference; Output candidate bounding boxes, target confidence scores, and corresponding feature embedding vectors; Determine whether the target detection confidence score is lower than a preset threshold, which is used to characterize whether the target features have changed; When the target features change, the feature embedding vector of the candidate target is normalized and similarity matching is performed with the category prototype. The confidence score of target detection is enhanced based on prototype matching similarity, and an enhanced confidence score is generated. Target screening and nonmaximum suppression are performed based on the enhanced confidence level to obtain the detection results; Based on the target samples that meet the conditions in the detection results, the corresponding category prototypes are updated.

[0032] An object detection system that adaptively evolves on an inference platform includes: The onboard inference and detection module achieves initial target screening and multi-dimensional feature output through the following steps: Image preprocessing: The input image is scaled proportionally and its boundaries are padded, and then normalized and converted to match the dimensions of the model input. Multi-scale feature extraction: Utilizes a deep convolutional network to perform forward inference and obtains original output tensors of different sizes; Dimension Reshaping and Parameter Parsing: The original tensor is reshaped into the (B, na, ny, nx, no) format to restore the anchor box dimensions; Candidate box decoding: Apply the Sigmoid activation function to the output parameters.

[0033] Perform coordinate transformation using preset grid offset and anchor frame parameters to calculate the center coordinates (xy) and width / height (wh) of the target candidate box: Feature separation and extraction: Simultaneously segment the decoded tensor to obtain the target existence confidence, class probability distribution, and fixed-length feature embedding vector.

[0034] The prototype matching module performs correlation analysis between candidate targets and preset categories by measuring vector distances within the feature space. Feature vector normalization: Perform L2 norm normalization on the feature embedding vector and the preset class prototype vector corresponding to each candidate box, and project the vectors onto the unit hypersphere. Similarity matrix operation: By calculating the product of the normalized feature vector and the prototype matrix, a cosine similarity matrix describing the degree of similarity between the target and each category is obtained; Category attribute prediction: The maximum value of each row in the similarity matrix (i.e., corresponding to each candidate box) is retrieved as the maximum similarity score, and its corresponding category index is mapped to the pre-detection category label of the target.

[0035] The confidence enhancement module applies a weighted correction to the detection confidence based on feature matching consistency. Match validity determination: Perform a quantitative comparison between the maximum similarity score and the preset similarity threshold, and filter out low-relevance feature interference through a step function or threshold mask; Confidence fusion calculation: The original target existence confidence score is multiplied by the matching validity determination result. Confidence reconstruction: This step enhances or corrects the initial detection confidence through similarity constraints, effectively suppressing spurious targets with mismatched class features.

[0036] The result filtering module outputs the final detection result through a multi-level filtering mechanism. Initial screening based on confidence level: Based on the enhanced confidence level, candidate targets with confidence levels below a preset threshold are eliminated; Non-maximum suppression (NMS): Residual candidate boxes are sorted in descending order of confidence, and redundant boxes with an overlap of higher than the suppression threshold are removed by iteratively calculating the overlap area. Geometric space mapping: The coordinates of the filtered candidate boxes are inversely scaled from the model inference space according to the preprocessing ratio, and boundary constraint clipping is performed to restore them to the original image physical coordinate system; Result encapsulation output: The output includes the final detection message containing the target category index, fusion confidence score, and original image coordinate information.

[0037] The prototype update module dynamically maintains the prototype based on detection feedback to address drift in target features. High-confidence feature sampling: Selecting target feature vectors that meet the high-confidence threshold and have correct spatial matching from the detection results; Feature centroid calculation: Calculate the mean of high-confidence feature vectors belonging to the same category within the current batch to obtain the instantaneous feature representation of that category; Exponential Moving Average Update: Performs an exponential moving average (EMA) update on the predefined category prototypes using the momentum factor. Statistical counting synchronization: Accumulate the sample observation counts for this category to achieve robust tracking of the evolution of target features under specific environments.

[0038] Beneficial effects: This invention provides a target detection method suitable for onboard inference platforms, enabling the detection model to complete the target detection task normally on the board. Furthermore, after the characteristics of the target being tested change (such as damage or occlusion), the target detection performance is dynamically improved during the inference stage by introducing a prototype matching mechanism, without the need to retrain the detection model.

[0039] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A target detection method that adaptively evolves on an inference platform, characterized in that, The target detection method includes: The onboard device acquires the image to be detected and inputs it into the target detection model for inference, performing initial target screening and multi-dimensional feature output; When the target features change, the feature embedding vector of the candidate target is normalized and similarity matching is performed with the category prototype. The confidence score of target detection is enhanced based on prototype matching similarity, and an enhanced confidence score is generated. Target screening and nonmaximum suppression are performed based on the enhanced confidence level to obtain the detection results; Based on the target samples that meet the conditions in the detection results, the corresponding category prototypes are updated.

2. The target detection method for adaptive evolution on an inference platform according to claim 1, characterized in that, The process of acquiring the image to be detected on the onboard device, inputting it into the target detection model for inference, and performing initial target screening and multi-dimensional feature output specifically includes: The image to be detected is acquired on the onboard device and input into the target detection model for inference; Output candidate bounding boxes, target confidence scores, and corresponding feature embedding vectors; Determine whether the target detection confidence level is lower than a preset threshold, which is used to characterize whether the target features have changed.

3. The target detection method for adaptive evolution on an inference platform according to claim 1, characterized in that, The specific steps of initial target screening and multi-dimensional feature output include: Image preprocessing: The input image is scaled proportionally and its boundaries are padded, and then normalized and converted to match the dimensions of the model input. Multi-scale feature extraction: Forward inference is performed using a deep convolutional network to obtain raw output tensors with multiple sizes; Dimension Reshaping and Parameter Parsing: Reshape the original tensor into (B, na, ny, nx, no) format to restore the anchor box dimensions; Candidate box decoding: Apply the Sigmoid activation function to the output parameters; Perform coordinate transformation using preset grid offset and anchor frame parameters to calculate the center coordinates (xy) and width / height (wh) of the target candidate box: Feature separation and extraction: Simultaneously segment the decoded tensor to obtain the target existence confidence, class probability distribution, and fixed-length feature embedding vector.

4. The target detection method for adaptive evolution on an inference platform according to claim 1, characterized in that, The process of normalizing the feature embedding vector of the candidate target and performing similarity matching with the category prototype when the target features change specifically includes: Feature vector normalization: Perform L2 norm normalization on the feature embedding vector and the preset class prototype vector corresponding to each candidate box, and project the vectors onto the unit hypersphere. Similarity matrix operation: By calculating the product of the normalized feature vector and the prototype matrix, a cosine similarity matrix describing the degree of similarity between the target and each category is obtained; Category attribute prediction: The maximum value of each row in the similarity matrix is ​​retrieved as the maximum similarity score, and the corresponding category index is mapped to the pre-detection category label of the target.

5. The target detection method for adaptive evolution on an inference platform according to claim 1, characterized in that, The enhancement of target detection confidence based on prototype matching similarity, specifically generating the enhanced confidence score, includes: Match validity determination: Perform a quantitative comparison between the maximum similarity score and the preset similarity threshold, and filter out low-relevance feature interference through a step function or threshold mask; Confidence fusion calculation: The original target existence confidence score is multiplied by the matching validity determination result. Confidence reconstruction: The initial detection confidence is enhanced or corrected by similarity constraints to suppress spurious targets with mismatched class features.

6. The target detection method for adaptive evolution on an inference platform according to claim 1, characterized in that, The target screening and non-maximum suppression performed based on the enhanced confidence level to obtain the detection results specifically include: Initial screening based on confidence level: Based on the enhanced confidence level, candidate targets with confidence levels below a preset threshold are eliminated; Non-maximum suppression: Residual candidate boxes are sorted in descending order of confidence, and redundant boxes with an overlap of higher than the suppression threshold are removed by iteratively calculating the overlap area. Geometric space mapping: The coordinates of the filtered candidate boxes are inversely scaled from the model inference space according to the preprocessing ratio, and boundary constraint clipping is performed to restore them to the original image physical coordinate system; Result encapsulation output: The output includes the final detection message containing the target category index, fusion confidence score, and original image coordinate information.

7. The target detection method for adaptive evolution on an inference platform according to claim 1, characterized in that, The step of updating the corresponding category prototype based on target samples that meet the conditions in the detection results specifically includes: High-confidence feature sampling: Selecting target feature vectors that meet the high-confidence threshold and have correct spatial matching from the detection results; Feature centroid calculation: Calculate the mean of high-confidence feature vectors belonging to the same category within the current batch to obtain the instantaneous feature representation of the corresponding category; Exponential Moving Average Update: Performs an exponential moving average update on the preset category prototypes using the momentum factor. Statistical counting synchronization: Accumulates the sample observation counts for the corresponding category, enabling robust tracking of target feature evolution under specific environments.

8. A target detection system that adaptively evolves on an inference platform, employing the target detection method that adaptively evolves on an inference platform as described in any one of claims 1-7, characterized in that, The target detection system includes: The onboard inference detection module is used to acquire the image to be detected on the onboard device, input it into the target detection model for inference, and perform initial screening of targets and output multi-dimensional features. The prototype matching module is used to normalize the feature embedding vector of the candidate target when the target features change, and to perform similarity matching with the category prototype. The confidence enhancement module is used to enhance the confidence of target detection based on prototype matching similarity and generate the enhanced confidence score. The results filtering module is used to perform target filtering and non-maximum suppression based on the enhanced confidence level to obtain the detection results. The prototype update module is used to update the corresponding category prototype based on the target samples that meet the conditions in the detection results.

Images