A target detection method, system and device based on hierarchical collaborative reasoning

Abstract

The application provides a target detection method, system and device based on hierarchical collaborative reasoning, relates to the technical field of target detection, and comprises the following steps: acquiring a target detection image for preliminary detection to obtain a preliminary detection result comprising boundary box coordinates, a category label and a confidence score; screening the preliminary detection result according to a confidence threshold to obtain a high-confidence target result and a target result to be reasoned; performing hierarchical collaborative reasoning on each target to be reasoned in the target result to be reasoned to obtain supporting evidence; and marking a supporting confidence target result based on the supporting evidence; and collecting the high-confidence result and the supporting confidence target result to generate a final target detection result. The application solves the problems of missed detection, false detection and semantic discontinuity caused by low confidence in a complex scene, and greatly improves the accuracy, robustness and reliability of target detection in a complex scene.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a target detection method, system and device based on hierarchical collaborative reasoning. Background Technology

[0002] Object detection, a core task in computer vision, aims to classify and locate objects in images or video sequences, serving as a prerequisite for intelligent perception, situational awareness, and decision-making control. In recent years, single-stage depth detectors, represented by the YOLO (YouOnly Look Once) series, have been widely applied in high-value scenarios such as target reconnaissance, security monitoring, autonomous driving, and industrial quality inspection due to their advantages of high speed and accuracy based on deep learning-based object detection algorithms.

[0003] However, with the increasing complexity of application environments, this depth detector has revealed the following shortcomings in real open environments: (1) Confidence bottleneck. When imaging conditions are poor (such as low light, backlight, rain, fog, motion blur), the target is partially occluded or the target size is extremely small, the signal-to-noise ratio of convolutional features decreases significantly, and the detection model will give a low confidence level, resulting in a large number of missed detections and false detections, causing the loss of effective targets.

[0004] (2) Semantic gaps. Existing detection models lack human-interpretable "common sense" and logical reasoning abilities. For some sets of objects with clear functions, spatial or compositional relationships, they can only score individual candidate boxes in isolation and cannot use contextual semantic consistency to perform cross-instance verification of low-confidence targets, which leads to systematic missed detections and category confusion in highly dynamic and complex scenarios where targets are mutually occluded.

[0005] (3) Poor domain generalization. General detection models require a large amount of labeled data and have a long training cycle, making it difficult to quickly adapt to complex scenarios with strong domain constraints and sensitive samples, such as national defense and security; thus, they cannot effectively utilize structured and systematic domain knowledge to improve the robustness and reliability of the model.

[0006] In summary, these problems are particularly prominent in complex scenarios such as national defense and security, where accuracy and reliability are extremely critical. Currently, there is no mature solution that deeply integrates the real-time capabilities of efficient detectors like YOLO with the explicit reasoning capabilities of knowledge graphs to address the issues of missed detections, false detections, and semantic gaps caused by low confidence levels in complex scenarios. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides a target detection method, system, and apparatus based on hierarchical collaborative reasoning, which solves the problems of missed detection, false detection, and semantic gaps caused by low confidence in complex scenarios, thereby significantly improving the accuracy, robustness, and reliability of target detection in complex scenarios.

[0008] In a first aspect, the present invention provides a target detection method based on hierarchical collaborative reasoning, the method comprising: The target detection image is acquired for preliminary detection, and preliminary detection results are obtained, including bounding box coordinates, class labels and confidence scores. The preliminary detection results are filtered based on the confidence threshold to obtain high-confidence target results and target results to be inferred; Hierarchical collaborative reasoning is performed on each target in the target results to be reasoned to obtain supporting evidence; and the support confidence target results are obtained based on the supporting evidence labeling; The high-confidence results and the supporting-confidence target results are combined to generate the final target detection result.

[0009] Furthermore, the hierarchical collaborative reasoning includes: Based on the category labels, query the current domain knowledge graph to obtain the entity relationships associated with the current inference target; In the spatial layer, the spatial relationships of the current target to be reasoned about are analyzed. If the spatial relationships are consistent with the entity relationships, they are recorded as supporting evidence. In the semantic layer, the semantic relationships of the current target to be reasoned about are analyzed. If the semantic relationships are consistent with the entity relationships, they are recorded as supporting evidence. If supporting evidence exists, the confidence increment of the current target to be inferred is calculated based on the entity relationship weight and the number of pieces of evidence. The confidence score of the current target to be inferred is updated based on the confidence increment, and it is determined whether it exceeds the confidence threshold. If it exceeds the threshold, the current target to be inferred is confirmed as a result supporting the confidence target; otherwise, the current target to be inferred is discarded.

[0010] Furthermore, the confidence increment of the current target to be inferred. The calculation method is as follows: ; in, This indicates the relationship weights set according to the actual scenario; Indicates the first The confidence level of each supporting piece of evidence. ; The number of pieces of evidence that provide supporting evidence.

[0011] Furthermore, the confidence score of the current target to be inferred is updated based on the confidence increment, as follows: ; in, This represents the updated confidence score of the current target to be inferred. This represents the confidence score of the target to be inferred. This represents the confidence increment.

[0012] Furthermore, the hierarchical collaborative reasoning also includes: Based on the current domain knowledge graph, the exclusion relationship is defined, and the confidence reduction of the current target to be inferred is calculated according to the exclusion relationship weight and the number of exclusion relationships. The confidence score of the current target to be inferred is updated based on the confidence increment and the confidence decrement, and it is determined whether it exceeds the confidence threshold.

[0013] Furthermore, the confidence score is: or .

[0014] Furthermore, if the updated confidence score still falls between the second confidence threshold and the first confidence threshold, it is sent to the next round of hierarchical collaborative inference and a loop strategy is executed to avoid infinite loops.

[0015] Furthermore, the preliminary detection results are filtered based on a confidence threshold, including: Receive preliminary detection results and obtain all target detection images and their confidence scores from the preliminary detection results; Set a first confidence threshold and a second confidence threshold, and compare the confidence score with the first confidence threshold and the second confidence threshold; Preliminary detection results that are higher than the first confidence threshold are directly marked as the high-confidence target results; Preliminary detection results that are below the second confidence threshold are discarded directly; Preliminary detection results that fall between the first confidence threshold and the second confidence threshold are marked as target results to be inferred.

[0016] In a second aspect, the present invention provides a target detection system based on hierarchical collaborative reasoning, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0017] Thirdly, the present invention provides a target detection device based on hierarchical collaborative reasoning, comprising a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0018] In summary, this invention provides a target detection method, system, and apparatus based on hierarchical collaborative reasoning. Compared with existing technologies, the technical solution conceived in this invention can achieve the following beneficial effects: (1) The present invention filters the preliminary detection results to obtain high-confidence target results and target results to be inferred; performs hierarchical collaborative reasoning on each target to be inferred in the target results to obtain target results with supporting confidence; and summarizes the high-confidence results and target results with supporting confidence to generate the final target detection results; the present invention deeply integrates the YOLO high-speed target detection results with the domain knowledge graph, and uses the domain knowledge graph to perform hierarchical collaborative reasoning to solve the problems of missed detection, false detection and semantic fragmentation caused by low confidence in complex scenarios such as target occlusion and blurring. It can not only effectively correct false detection and complete missed detection, but also reduce the sensitivity of the model to factors such as image quality and illumination changes, and greatly improve the accuracy, robustness and reliability of low-confidence target detection image judgment.

[0019] (2) This invention proposes a dynamic confidence adjustment mechanism based on contextual logical relationships. Through the dual consistency test of spatial layer and semantic layer and the hierarchical collaborative reasoning of entity relationship weighting, the prior relationships in the domain knowledge graph are transformed into measurable supporting evidence by real-time querying of the domain knowledge graph and matching contextual evidence. A reasonable confidence increment is calculated for low confidence targets, and then the confidence is dynamically updated. This transforms low confidence targets that have been misjudged or abandoned by YOLO back into reliable detection results. It significantly improves the detection accuracy, retrieves low confidence targets that have been suppressed due to occlusion, ambiguity, or extremely small scale, solves the problem of judging low confidence targets, and reduces the false negative rate and false positive rate.

[0020] (3) On the one hand, the present invention dynamically updates the confidence level through dual consistency verification of spatial layer and semantic layer, so that low confidence targets can obtain increased confidence when satisfying context logic, significantly reducing missed detections in complex occlusion and blurry scenes. At the same time, the entire inference chain is traceable, improving accuracy and interpretability. On the other hand, the invention introduces exclusion relationship and confidence reduction, forming a two-way dynamic adjustment with confidence increment. This can prevent the abuse of false evidence and allow negative evidence to lower the level of suspicious targets in a timely manner. It can adapt to changes in illumination and viewpoint without additional training data, which greatly improves the accuracy, robustness and reliability of target detection image judgment.

[0021] (4) This invention uses a dual threshold and intermediate band screening strategy to clearly classify the preliminary detection results into three categories: high confidence, pending reasoning, and direct discard. This eliminates obvious false detections and focuses on potential targets in the early stages of operation, reduces the computational load of subsequent domain knowledge graph reasoning, ensures real-time detection, provides accurate input for hierarchical collaborative reasoning, avoids noise interference, and improves the accuracy of confidence increment.

[0022] (5) Without increasing the complexity of the model, the present invention significantly improves the detection accuracy and robustness, and the reasoning process is transparent and traceable, which enhances the interpretability of the results. It not only meets the requirements of high-risk fields such as autonomous driving and security for algorithm interpretability.

[0023] (6) This invention utilizes the YOLO basic detector and domain knowledge base, which can be replaced and customized according to different application fields. It has strong versatility and good scalability and adaptability. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the method steps of a target detection method, system and device based on hierarchical collaborative reasoning provided by the present invention; Figure 2 This is a schematic diagram illustrating the principle of a target detection method, system, and device based on hierarchical collaborative reasoning provided by the present invention. Figure 3 This is a schematic diagram of the system structure of a target detection method, system and device based on hierarchical collaborative reasoning provided by the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0027] It should be noted that, in the description of the embodiments of the present invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a method, step, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to the method, step, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the method, step, or apparatus that includes that element.

[0028] Firstly, this invention proposes a target detection method, system, and device based on hierarchical collaborative reasoning. By introducing a domain knowledge graph, it verifies, corrects, and completes the preliminary results of detection algorithms such as YOLO. It utilizes spatial and semantic relationships to perform logical reasoning on low-confidence detected targets, solving the problems of missed detections, false detections, and semantic gaps caused by low confidence in complex scenarios, thereby improving the accuracy, robustness, and reliability of image judgment for low-confidence target detection.

[0029] Specifically, such as Figure 1 and Figure 2 As shown, the method includes: S100: Acquire the target detection image for preliminary detection, and obtain preliminary detection results including bounding box coordinates, class labels, and confidence scores.

[0030] Specifically, a basic detector based on the YOLO series of algorithms can be used for initial detection. The YOLO series includes YOLOv5, YOLOv7, and YOLOv8. It should be noted that this basic detector uses the YOLO series of algorithms to obtain information including bounding box coordinates and category labels. and confidence score The preliminary test results are based on relatively conventional existing technologies, which will not be elaborated on here.

[0031] S200: Filter the preliminary detection results according to the confidence threshold to obtain high-confidence target results and target results to be inferred.

[0032] As an example, the preliminary detection results are screened based on a confidence threshold, including: Receive preliminary detection results and obtain all target detection images and their confidence scores from the preliminary detection results; Set a first confidence threshold and a second confidence threshold, and compare the confidence score with the first confidence threshold and the second confidence threshold; Preliminary detection results that exceed the first confidence threshold are directly marked as high-confidence target results; Preliminary detection results that are below the second confidence threshold are discarded directly; Preliminary detection results that fall between the first confidence threshold and the second confidence threshold are marked as target results to be inferred.

[0033] The first confidence threshold is also known as the high confidence threshold, and the second confidence threshold is also known as the low confidence threshold. For example, setting the first confidence threshold... for Second confidence threshold for The confidence score of the current target detection image is: ,like If it is lower, it is directly marked as a high-confidence target result; if it is lower, it is marked as a high-confidence target result. If it is, then discard it directly; if If it is, then it is marked as the target result to be inferred.

[0034] This invention employs a dual-threshold and intermediate-band screening strategy, which clearly categorizes preliminary detection results into three types: high confidence, pending reasoning, and direct discard. This approach eliminates obvious false detections and focuses on potential targets in the early stages of operation, reducing the computational load of subsequent domain knowledge graph reasoning, ensuring real-time detection, providing accurate input for hierarchical collaborative reasoning, avoiding noise interference, and improving the accuracy of confidence increments.

[0035] S300: Perform hierarchical collaborative reasoning on each target to be reasoned in the result of the reasoning target to obtain supporting evidence; and obtain the target result of support confidence based on the supporting evidence label.

[0036] As one example, hierarchical collaborative reasoning includes: Based on category labels, query the current domain knowledge graph to obtain the entity relationships associated with the current reasoning target; In the spatial layer, analyze the spatial relationships of the current target to be reasoned about. If the spatial relationships are consistent with the entity relationships, they are recorded as supporting evidence. In the semantic layer, analyze the semantic relationships of the current target to be reasoned about. If the semantic relationships are consistent with the entity relationships, they are recorded as supporting evidence. If supporting evidence exists, the confidence increment of the current target to be inferred is calculated based on the entity relationship weight and the number of pieces of evidence. The confidence score of the current target to be inferred is updated based on the confidence increment, and it is determined whether it exceeds the confidence threshold. If it exceeds the threshold, the current target to be inferred is confirmed as a result that supports the confidence target; otherwise, the current target to be inferred is discarded.

[0037] In other words, this invention transforms prior relations in the domain knowledge graph into measurable supporting evidence through a dual consistency check of the spatial layer and semantic layer and hierarchical collaborative reasoning with entity relation weighting. It calculates a reasonable confidence increment for low-confidence targets and dynamically updates the confidence, thereby converting low-confidence targets that were misjudged or abandoned by YOLO back into reliable detection results. This significantly improves detection accuracy, retrieves low-confidence targets that were suppressed due to occlusion, ambiguity, or extremely small scales, solves the problem of accurate judgment of low-confidence targets, and reduces the false negative and false positive rates.

[0038] It should be noted that the current domain knowledge graph includes entities, relations, and attributes; among them, relations include spatial relations and semantic relations.

[0039] Specifically, the establishment of the current domain knowledge graph includes: defining a three-layer model of entities, relations, and attributes based on the top-level ontology, and setting inverse constraints on relations; extracting triples from structured text, semi-structured text, and unstructured text in sequence and validating them; and storing the validated triples in a hybrid form of attribute graph and RDF to complete the establishment of the current domain knowledge graph.

[0040] In other words, for specific application scenarios such as cockpits, transportation scenarios, and industrial workshops, we define relevant entities, relationships, and attributes to form a domain knowledge graph.

[0041] Entities correspond to target categories, such as "steering wheel," "seat," "pedestrian," and "traffic light." Relationships define the spatial and semantic relationships between entities, such as "A is Part Of B," meaning A is part of B; "A is Located In B," meaning A is located inside B; "A often Co Occurs With B," meaning A usually coexists with B; "A spatially Above B," meaning A is above B, etc. Attributes are the properties that entities can possess, such as "typical size" and "common color," used to assist in reasoning.

[0042] Furthermore, taking the cockpit scenario as an example, the specific steps for extracting spatial and semantic relationships from the current domain knowledge base are as follows: First, query the Neo4j attribute graph for high-confidence entities in the same frame as the target to be inferred, call the preset geometric functions to calculate the center point offset, IoU, and azimuth angle, and automatically generate spatial relationships such as "spatially Above / Below"; then, perform cosine matching between the RoI feature vector of the target to be inferred and the embedding representation of each entity in the knowledge base RDF to determine the most likely category, and then extract the "is Part Of / often Co Occurs With" function or co-occurrence relationship by traversing the graph in one hop, and output the semantic relationship after aggregation by the graph attention network, thereby completing the synchronous extraction of spatial and semantic relationships.

[0043] By querying the current domain knowledge graph based on category labels, we can obtain the entity relationships associated with the current reasoning target. This involves querying the knowledge base for other entities with strong relationships to the target based on its category labels. Specifically, this includes: Using the category label of the current target to be reasoned (such as "mobile phone") as the key, locate the corresponding entity node in the index of the current domain knowledge graph; Traverse outwards from the node in one hop, extracting all edges of predefined types and their connected entities to obtain the entity relationships of triples; for example, “mobile phone—often Co Occurs With—coffee cup”, “mobile phone—spatiallyAbove—seatbelt slot”, etc.

[0044] Finally, after filtering by edge weight, a set of entity relationships directly related to the current inference target is formed for subsequent consistency verification at the spatial and semantic layers.

[0045] At the spatial layer, the spatial relationships of the target to be inferred are analyzed. If the spatial relationship matches the entity relationship, it is recorded as supporting evidence; for example, "A is located in B," "A spatially above B," etc. If the spatial relationship in the target image matches the entity relationship queried in the current domain knowledge graph, it is recorded as one piece of supporting evidence. At the semantic layer, the semantic relationships of the target to be inferred are analyzed. If the semantic relationship matches the entity relationship, it is recorded as supporting evidence; for example, "A is part of B," "A often co occurs with B," etc. If the semantic relationship in the target image matches the entity relationship queried in the current domain knowledge graph, it is recorded as one piece of supporting evidence.

[0046] In other words, based on entity relationships, we check whether the confirmed targets with high confidence in the results of the reasoning target exist in the spatial and semantic layers.

[0047] For example, at the spatial layer, suppose we detect a target aircraft 1 with low confidence that requires inference, and a high-confidence helipad, with the aircraft positioned directly above the helipad in the image. Querying the current domain knowledge graph reveals a spatial relationship: the aircraft is on the helipad. Therefore, the high-confidence helipad provides one piece of supporting evidence for the low-confidence target aircraft 1. Similarly, we analyze all high-confidence confirmed targets in the image that have a spatial relationship with target aircraft 1, match them with the current domain knowledge graph, and calculate the number of supporting pieces of evidence.

[0048] At the semantic layer, suppose we detect target aircraft 1 with low confidence requiring inference, and target aircraft 2 with high confidence. Querying the current domain knowledge graph reveals that target aircraft 1 and target aircraft 2 are of the same type and typically appear at the same airport. Therefore, the high-confidence target aircraft 2 provides one piece of supporting evidence for the low-confidence target aircraft 1 requiring inference. Similarly, we analyze all high-confidence confirmed targets in the image that have a semantic relationship with target aircraft 1, match them with the current domain knowledge graph, and calculate the number of supporting pieces of evidence.

[0049] Through rich spatial and semantic relationships, it is possible to understand and utilize deep domain knowledge that goes far beyond visual representation, significantly improving the algorithm's cognitive ability in professional scenarios. This reasoning process is transparent and traceable, and can provide a basis for decision-making, such as believing that B exists because A was detected, thus meeting the requirements for algorithm interpretability in some high-risk fields.

[0050] As an example, the confidence increment of the current target to be inferred. The calculation method is as follows: ; in, This indicates the relationship weights set according to the actual scenario; Indicates the first The confidence level of each supporting piece of evidence. ; The number of supporting evidences includes the number of evidences at the spatial level and the number of evidences at the semantic level.

[0051] It should be noted that relation weights are set according to the actual scenario. For example, the relation "A often Co Occurs With B" should have its weight defined based on the probability of coexistence and non-coexistence, while the relation "A is Part Of B" should have a higher weight. Similarly, before application, it is necessary to sort out the spatial and semantic relationships according to the actual application scenario and quantify the weights through empirical values ​​or other methods.

[0052] Furthermore, the confidence score of the current target to be inferred is updated based on the confidence increment, as follows: ;in, This represents the updated confidence score of the current target to be inferred. This represents the confidence score of the target to be inferred. This represents the confidence increment.

[0053] As one example, hierarchical collaborative reasoning also includes: Based on the current domain knowledge graph, exclusion relations are defined. The confidence reduction of the current target to be reasoned about is calculated according to the weight and number of exclusion relations. The confidence score of the current target to be reasoned about is updated based on the confidence increment and confidence reduction, and it is determined whether it exceeds the confidence threshold. It should be noted that exclusion relations can be used for negative reasoning to reduce the confidence of targets with exclusionary evidence.

[0054] As an example, the confidence level of the target to be inferred is reduced. The calculation method is as follows: ;in, This indicates the exclusion relationship weights set according to the actual scenario; Indicates the first Confidence level of an exclusion relationship ; Indicates the number of exclusion relationships.

[0055] At this point, the confidence score can be: or .

[0056] It should also be noted that if the updated confidence score still falls between the second confidence threshold and the first confidence threshold, that is, it is still the target result to be inferred, then it will be sent to the next round of hierarchical collaborative inference and a loop strategy will be executed to avoid infinite loops.

[0057] Furthermore, the cyclical strategy includes one or more of the following: expanding the search scope, enabling conflict resolution arbitration, and manual review.

[0058] Expand the search scope. Specifically, at the spatial layer, relax the query radius / IoU threshold setting (e.g., 20%) to introduce more surrounding entities to participate in the relationship; at the semantic layer, allow jumping one level outward in the knowledge graph (e.g., 2-hop) to mine indirect evidence, such as mobile phone → charging cable → USB port, etc.

[0059] Expanding the scope of evidence search, specifically, recording supporting / rejecting evidence from the current and previous rounds. If the evidence set remains unchanged for two consecutive rounds and the confidence level change is <0.02, it is considered evidence exhaustion, and the loop stops. The number of supporting evidence is compared to the number of rejecting relationships. If the number of supporting evidence is greater than the number of rejecting relationships, only the number of supporting evidence is used for calculation. If the number of exclusionary relations is greater than the number of supporting evidence, then only the number of exclusionary relations is used for calculation, that is, .

[0060] Manual review, specifically, involves the initial detection results that cannot be determined even after expanding the search scope and enabling conflict resolution arbitration. These results are then marked with a review tag and returned to the background for manual review.

[0061] By using a loop strategy to prevent infinite recursion, truly difficult boundary cases are transformed into continuously improving data, thus ensuring the real-time performance and robustness of object detection.

[0062] For example, in the domain knowledge graph, an exclusion relation is defined as: mobile phone — mutually Exclusive With — steering wheel, with a weight w = 0.4, indicating that the probability of both appearing simultaneously in the driver's hands area is extremely low. Preliminary detection results are: the confidence score for the steering wheel is 0.92 (high confidence, already retained), and the confidence score for the mobile phone is 0.45 (to be inferred). When an exclusion relation is triggered, the confidence score is reduced; only one exclusion relation is effective. If there is no other supporting evidence, the confidence score for the updated phone is [score to be inserted]. If the value is less than the threshold of 0.3, the "phone" is marked as a false detection and discarded, thus avoiding incorrectly labeling normal driving scenarios where the driver is holding the steering wheel as "distracted phone use." If there is other supporting evidence regarding the phone... The confidence score after the phone update is: If the result is greater than the threshold of 0.3 but less than the threshold of 0.7, the "phone" is still marked as the target result to be inferred, and is sent to the next round of hierarchical collaborative reasoning, and the loop strategy is executed.

[0063] S400: Summarizes high-confidence results and supporting-confidence target results to generate the final target detection result.

[0064] For example, high-confidence results and results supporting the target confidence can be written into the same result table, and after deduplication and confidence calibration, the final target detection results can be output in the required format.

[0065] In a second aspect, the present invention also provides a target detection system based on hierarchical collaborative reasoning, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0066] Furthermore, such as Figure 3 As shown, the processor may include an object detection module, a domain knowledge base, a hierarchical collaborative reasoning module, and a result fusion module.

[0067] Thirdly, the present invention also provides a target detection device based on hierarchical collaborative reasoning, comprising a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0068] The specific technical features of the system and product are consistent with the technical features of the method, and will not be elaborated here.

[0069] In summary, this invention obtains preliminary detection results based on efficient object detection algorithms such as YOLO, and is supported by a domain knowledge graph. Utilizing the semantic and spatial relationships between entities within the pre-constructed domain knowledge graph, a hierarchical collaborative reasoning engine performs logical reasoning and verification on low-confidence detection targets. It also performs logical verification and adaptive updates to the confidence level of the detection results, achieving an effective fusion of data-driven and knowledge-driven approaches. This not only allows for differentiated application of knowledge reasoning to detection results with varying confidence levels, effectively correcting false detections, supplementing missed detections, and improving the detection confidence of key targets, but also integrates structured prior knowledge with data-driven perception. This enhances the system's adaptability to harsh environments such as occlusion and blurring, significantly improving the accuracy, robustness, and reliability of object detection in complex scenarios, achieving highly robust and interpretable intelligent visual perception.

[0070] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0071] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0072] In the several embodiments provided in this application, it should be understood that the disclosed methods or systems can be implemented in other ways. For example, the embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0073] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of embodiments of this disclosure upon considering the specification and practicing the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

[0074] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0075] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements 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 based on hierarchical collaborative reasoning, characterized in that, The method includes: The target detection image is acquired for preliminary detection, and preliminary detection results are obtained, including bounding box coordinates, class labels and confidence scores. The preliminary detection results are filtered according to a confidence threshold to obtain high-confidence target results and target results to be inferred; if If it is, then it is marked as the target result to be reasoned, where, This represents the first confidence threshold. This represents the second confidence threshold; For each target in the proposed reasoning results, hierarchical collaborative reasoning is performed. Based on the category label, the current domain knowledge graph is queried to obtain the entity relationship associated with the current target. The spatial relationship of the current target is analyzed at the spatial layer. If the spatial relationship is consistent with the entity relationship, it is recorded as supporting evidence. The semantic relationship of the current target is analyzed at the semantic layer. If the semantic relationship is consistent with the entity relationship, it is recorded as supporting evidence. Thus, supporting evidence is obtained. If supporting evidence exists, the confidence increment of the current target to be inferred is calculated based on the entity relation weight and the amount of evidence. The confidence score of the current target to be inferred is updated based on this confidence increment, and it is determined whether it exceeds a confidence threshold. If it does, the current target to be inferred is confirmed as supporting the confidence target result; otherwise, the current target to be inferred is discarded. The confidence increment of the current target to be inferred... The calculation method is as follows: ; This indicates the relationship weights set according to the actual scenario; Indicates the first The confidence level of each supporting piece of evidence. ; The number of pieces of evidence indicating supporting evidence; The high-confidence results and the supporting-confidence target results are combined to generate the final target detection result.

2. The target detection method based on hierarchical collaborative reasoning according to claim 1, characterized in that, The confidence score of the current target to be inferred is updated based on the confidence increment as follows: ; in, This represents the updated confidence score of the current target to be inferred. This represents the confidence score of the target to be inferred. This represents the confidence increment.

3. The target detection method based on hierarchical collaborative reasoning according to claim 1, characterized in that, The hierarchical collaborative reasoning also includes: Based on the current domain knowledge graph, the exclusion relationship is defined, and the confidence reduction of the current target to be inferred is calculated according to the exclusion relationship weight and the number of exclusion relationships. The confidence score of the current target to be inferred is updated based on the confidence increment and the confidence decrement, and it is determined whether it exceeds the confidence threshold.

4. The target detection method based on hierarchical collaborative reasoning according to claim 3, characterized in that, The confidence score is: or ; This indicates a reduction in the confidence level of the target to be inferred.

5. The target detection method based on hierarchical collaborative reasoning according to claim 3, characterized in that, If the updated confidence score still falls between the second confidence threshold and the first confidence threshold, it is sent to the next round of hierarchical collaborative inference and a loop strategy is executed to avoid infinite loops.

6. The target detection method based on hierarchical collaborative reasoning according to claim 1, characterized in that, The preliminary detection results are filtered based on a confidence threshold, including: Receive preliminary detection results and obtain all target detection images and their confidence scores from the preliminary detection results; Set a first confidence threshold and a second confidence threshold, and compare the confidence score with the first confidence threshold and the second confidence threshold; Preliminary detection results that are higher than the first confidence threshold are directly marked as the high-confidence target results; Preliminary detection results that are below the second confidence threshold are discarded directly; Preliminary detection results that fall between the first confidence threshold and the second confidence threshold are marked as target results to be inferred.

7. A target detection system based on hierarchical collaborative reasoning, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

8. A target detection device based on hierarchical collaborative reasoning, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 6.

Images