Multi-model target detection method driven by influence factors in complex dynamic scene

Abstract

The invention relates to an influence factor-driven multi-model target detection method in a complex dynamic scene, which comprises the following steps of: systematically constructing a quantitative system covering multi-dimensional scene influence factors, and creating an evaluation data set with different factor combination labels according to the quantitative system; performing comprehensive performance evaluation on all candidate models in a pre-constructed lightweight model library by using the data sets so as to determine an optimal model for each typical influence factor combination, and establishing a mapping relation table; for a new target detection task, the scene of the new target detection task is analyzed in real time and quantized to obtain a specific influence factor combination, and then the optimal lightweight model can be quickly matched and loaded to execute detection by querying the mapping table. According to the method and the device, self-adaptive and low-delay model selection and calling of a complex and changeable scene on the end side equipment with limited resources are realized, and the overall performance and robustness of a target detection system in a changeable environment are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and target detection technology, specifically to a multi-model target detection method driven by influencing factors in complex dynamic scenarios. Background Technology

[0002] As one of the most challenging problems in the field of computer vision, object detection has been widely studied and explored by all sectors of society in recent years. The rapid development of deep learning technology has also greatly improved the accuracy of object detection, laying the foundation for its application in many scenarios such as face recognition, pedestrian detection, remote sensing object detection, and vehicle detection.

[0003] Current research on target detection primarily focuses on specific scenarios. However, once the scenario changes, the performance of detection models degrades, making it difficult to meet application requirements. For example, in target detection tasks using UAV remote sensing images, significant changes in the UAV's flight altitude, focal length, and shooting angle lead to substantial changes in the pixel resolution and pose of the target in the remote sensing image. Consequently, the features extracted by deep neural networks significantly impact the accuracy of target detection and recognition. Furthermore, in real-time detection scenarios, environmental factors such as clouds, rain, fog, and occlusion introduce interference noise into the remote sensing image, further reducing the performance of the detection model.

[0004] In summary, the existing technology has the following drawbacks:

[0005] (1) Existing methods focus on the characteristics of the image content itself (such as target scale and appearance), lacking joint modeling and analysis of systematic influencing factors such as imaging source and environment, resulting in an insufficient and in-depth portrayal of complex dynamic scenes, making it difficult to support accurate model selection decisions.

[0006] (2) Existing model selection strategies are mostly static or based on finite rules, which cannot dynamically and adaptively match the optimal model based on real-time perceived multi-dimensional and fine-grained influencing factors, and lack flexibility when facing unknown or complex influencing factor scenarios.

[0007] (3) Existing solutions usually require retraining the model for each sub-scenario or rely on large generalized models, failing to make full use of existing, diverse, lightweight model library resources, and failing to establish an efficient model performance pre-evaluation and mapping mechanism, resulting in low resource utilization and slow response speed.

[0008] (4) Under the condition of strictly limited edge resources, existing technologies rarely consider how to design the complex scenario analysis, model evaluation and selection process itself in a lightweight manner to ensure the real-time performance and low power consumption of the entire adaptive selection strategy. Summary of the Invention

[0009] In view of the above-mentioned technical problems, this invention proposes a multi-model target detection method driven by influencing factors in complex dynamic scenes. It can perform multi-dimensional quantitative analysis of influencing factors in complex dynamic scenes, establish a precise mapping relationship between scene factors and the optimal lightweight model through systematic evaluation, and then realize adaptive and real-time optimal model selection and invocation for new sensing scenes. Thus, while ensuring the feasibility of edge deployment, it significantly improves the overall performance and robustness of the target detection system in changing environments.

[0010] The technical solution to the technical problem of this invention is: a multi-model target detection method driven by influencing factors in complex dynamic scenes, comprising the following steps:

[0011] Step S1: Obtain or train multiple lightweight object detection models to build an initial model library;

[0012] Step S2: Decompose the complex dynamic scene that affects the target detection performance, identify multiple quantifiable scene influencing factors, and quantify and classify each influencing factor.

[0013] Step S3: Based on the scene influencing factor system after quantification and grading, the original image database is labeled to construct multiple evaluation datasets with different combinations of influencing factor labels;

[0014] Step S4: Systematically evaluate each target detection model in the initial model library using the multiple evaluation datasets, determine the optimal target detection model corresponding to each combination of influencing factors, and establish a mapping relationship table between the combination of influencing factors and the optimal target detection model;

[0015] Step S5: For a new target detection task, analyze its corresponding complex scenario, determine its specific combination of influencing factors under the scenario influencing factor system, query the mapping relationship table to select the corresponding optimal target detection model to execute the detection task.

[0016] According to one technical solution of the present invention, the scene influencing factors include at least three of the following: platform characteristics, load characteristics, environmental characteristics, target characteristics, and target camouflage characteristics.

[0017] According to one technical solution of the present invention, the platform characteristics include at least one of the following: type of platform, track height, attitude change, and vibration amplitude;

[0018] The payload characteristics include at least one of the following: sensor type, resolution, focal length, field of view, and imaging mode;

[0019] The environmental characteristics include at least one of the following: country / region, weather conditions, sea breeze, and sea state;

[0020] The target characteristics include at least one of the following: target size, target attitude, target category, and operating condition;

[0021] The target camouflage characteristics include at least one of optical camouflage and smoke concealment.

[0022] According to a technical solution of the present invention, the quantitative classification of each influencing factor includes:

[0023] For continuous influencing factors, their value range is divided into multiple levels, and a numerical range or descriptive label is set for each level;

[0024] For discrete influencing factors, their different values ​​are directly used as different levels.

[0025] According to one technical solution of the present invention, constructing multiple evaluation datasets with different combinations of influencing factor labels includes:

[0026] For each image in the original image database, based on its imaging conditions and content characteristics, the quantitative level of each influencing factor is labeled according to the scene influencing factor system to form a combination label of the influencing factors of the image;

[0027] Images are classified based on the combination of influencing factors labels, and images with the same or similar combination of influencing factors labels are grouped together to form multiple evaluation sub-datasets.

[0028] According to one technical solution of the present invention, the systematic evaluation of each target detection model in the initial model library includes:

[0029] Each evaluation subset is input into each object detection model to obtain the model's detection results;

[0030] For each model's detection results on each evaluation dataset, multiple evaluation metrics are calculated, including at least a detection accuracy metric and a computational efficiency metric.

[0031] A model performance matrix is ​​established based on evaluation metrics. The rows of the matrix represent different object detection models, the columns represent different evaluation datasets, and the matrix elements are the comprehensive evaluation scores of the models on the corresponding datasets.

[0032] According to one technical solution of the present invention, determining the optimal target detection model corresponding to each combination of influencing factors includes:

[0033] For each evaluation dataset corresponding to a specific combination of influencing factors, the target detection model with the highest comprehensive evaluation score in that column of the model performance matrix is ​​selected as the optimal target detection model for that combination of influencing factors.

[0034] When the combined evaluation scores of multiple models differ within a preset threshold, the model with higher computational efficiency is selected as the optimal target detection model.

[0035] According to one technical solution of the present invention, the establishment of the mapping relationship table between the combination of influencing factors and the optimal target detection model includes:

[0036] Construct a multidimensional lookup table, where the index of the table is the code of the combination of influencing factors, and the table entries are the identification information and performance parameters of the corresponding optimal target detection model;

[0037] The mapping table is stored in the local storage of the edge device or in a cloud server and is provided for querying via an API interface.

[0038] According to one technical solution of the present invention, the method further includes a mapping table update step:

[0039] Regularly collect new evaluation data and newly released object detection models;

[0040] Re-execute the model evaluation and mapping table construction steps to generate an updated mapping table;

[0041] The updated mapping table is pushed to the edge device using an incremental update method.

[0042] According to one technical solution of the present invention, when no perfectly matching combination of influencing factors is found in the mapping table, one of the following strategies is adopted:

[0043] Select the target detection model corresponding to the closest combination of influencing factors;

[0044] From target detection models corresponding to multiple similar combinations of influencing factors, model fusion is performed based on similarity weights;

[0045] The online evaluation module is activated to select multiple candidate models from the model library for rapid evaluation, determine the most suitable model, and update the mapping table.

[0046] Compared with the prior art, the present invention has the following advantages:

[0047] This invention provides a multi-model target detection method driven by influencing factors in complex dynamic scenarios. By systematically identifying, decomposing, and quantifying multi-dimensional scenario factors (such as platform, load, environment, target, and camouflage characteristics) that affect target detection performance, it achieves a refined and structured description of complex dynamic scenarios, laying a solid foundation for accurate model performance evaluation and matching, and overcoming the limitations of single-dimensional scenario characterization in existing technologies.

[0048] This invention establishes a highly reliable mapping table of "scene factors - optimal model"; during the detection phase, it is only necessary to quickly analyze the real-time scene and look up the table to determine the optimal model, thus achieving an effective balance between adaptive selection and the real-time requirements of the edge.

[0049] This invention fully utilizes and evaluates existing diverse lightweight model library resources, and matches the best-performing existing models for different scenario conditions through systematic benchmark testing. This avoids the huge overhead of retraining models for each sub-scenario, and significantly improves the utilization efficiency of model resources and system response speed.

[0050] The core operations of the entire selection strategy of this invention (scenario factor quantification and table lookup matching) can be designed as a lightweight process, which can be efficiently deployed on resource-constrained edge devices together with small models in the model library. This realizes end-to-end intelligentization from scene perception to model execution, and improves the overall adaptability and robustness of the system in changing environments. Attached Figure Description

[0051] Figure 1 A flowchart illustrating a multi-model target detection method driven by influencing factors in a complex dynamic scene according to an embodiment of the present invention;

[0052] Figure 2 This diagram illustrates the composition of a multi-model target detection system driven by influencing factors in a complex dynamic scene according to an embodiment of the present invention. Detailed Implementation

[0053] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0054] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The embodiments cannot be described in detail here, but the embodiments of the present invention are not limited to the following embodiments.

[0055] like Figure 1 As shown, the present invention provides a multi-model target detection method driven by influencing factors in complex dynamic scenes, comprising the following steps:

[0056] Step S1: Obtain or train multiple lightweight object detection models to build an initial model library;

[0057] Specifically, a series of object detection models optimized for edge deployment can be collected from publicly available model libraries or historical accumulations, such as YOLOv5n, YOLOv8s, MobileNetV3-SSD, and EfficientDet-Lite0. An initial candidate model library M = {m1, m2, ..., mk} is then constructed based on this large number of object detection models. The construction of the model library can be a one-time event or it can be updated periodically.

[0058] Step S2: Decompose the complex dynamic scene that affects the target detection performance, identify multiple quantifiable scene influencing factors, and quantify and classify each influencing factor.

[0059] For example, in drone inspection scenarios, key influencing factors identified may include: platform characteristics (such as flight speed and attitude angle fluctuations), payload characteristics (such as camera focal length and image resolution), environmental characteristics (such as clouds, rain, and fog), target characteristics (such as vehicle size and direction of movement), and target camouflage characteristics (such as camouflage). A quantitative grading standard is defined for each factor, such as classifying cloud cover into "no clouds," "light clouds," "medium clouds," and "heavy clouds."

[0060] Step S3: Based on the scene influencing factor system after quantification and grading, the original image database is labeled to construct multiple evaluation datasets with different combinations of influencing factor labels;

[0061] Obtain a large-scale raw image database D. For each image d in database D, based on its metadata and image content analysis, and according to the quantization system defined in step S2, label it with a set of influencing factor level labels d. i =(d i1 , d i2 , ..., d in ), where d ij This represents the ranking of the j-th influencing factor on that image. All labeled images constitute the labeled dataset D. D Then, based on the similarity or identicalness of the labels (e.g., allowing for minor differences), D is... D The model is divided into multiple subsets {D1, D2, ..., Dp}, each representing a typical combination of influencing factors, used for model performance evaluation.

[0062] Step S4: Systematically evaluate each target detection model in the initial model library using the multiple evaluation datasets, determine the optimal target detection model corresponding to each combination of influencing factors, and establish a mapping relationship table between the combination of influencing factors and the optimal target detection model;

[0063] For each evaluation dataset D j(Represents a combination of influencing factors), which is used for each model m in the model library M. i The evaluation yields a set of performance metrics, Score(i, j), such as mAP (mean accuracy), FPS (frames per second), and model size. A comprehensive scoring function can be designed. This combines multiple indicators into a single score, for example: F. , where α, β, γ are weighting coefficients. Through calculation, a performance matrix P is obtained, where... For each column j (i.e., each combination of influencing factors), find the model with the highest score in that column, which is the optimal model for that combination. Finally, store the correspondence between all combinations of influencing factors and the optimal model for that combination in a mapping table T.

[0064] This process can be designed as an offline process.

[0065] Step S5: For a new target detection task, analyze its corresponding complex scenario, determine its specific combination of influencing factors under the scenario influencing factor system, query the mapping relationship table to select the corresponding optimal target detection model to execute the detection task.

[0066] When the edge device acquires a new image to be detected, it first invokes a lightweight scene analysis module. This module analyzes the image and outputs its influencing factor level labels. Then, using these labels as indexes, it queries a locally stored mapping table T. The query can be designed as an exact match or a nearest neighbor match. The matching or closest combination of influencing factors is found, and its corresponding optimal model identifier is obtained. Finally, the corresponding optimal model is loaded from the local model library, and object detection is performed on the image to be detected, outputting the detection results.

[0067] The detection phase is conducted online.

[0068] In some embodiments of the present invention, the scene influencing factors include at least three of the following: platform characteristics, load characteristics, environmental characteristics, target characteristics, and target camouflage characteristics.

[0069] Platform characteristic quantification: This mainly refers to mobile platforms that carry imaging equipment. For example:

[0070] Vibration amplitude: Accelerometer data is acquired through the platform's built-in IMU (Inertial Measurement Unit), and its variance or peak value within a certain time window is calculated and classified into levels such as "stable", "mild vibration" and "severe vibration".

[0071] Attitude angles (pitch / roll): acquired via IMU, the angle variation range is divided into multiple intervals, such as "horizontal (<5°)", "small angle tilt (5°~15°)", and "large angle tilt (>15°)".

[0072] Payload characteristic quantification: mainly for the imaging sensor itself, for example:

[0073] Sensor type: discrete values, such as "visible light RGB camera", "thermal infrared camera", "multispectral camera".

[0074] Equivalent focal length: A continuous value that can be quantified into levels such as "wide-angle", "standard", and "telephoto", based on the size of the field of view.

[0075] Imaging resolution: discrete values, such as "720p", "1080p", "4K".

[0076] Environmental characteristic quantification: mainly refers to the external conditions during imaging, such as:

[0077] Country / Region: Obtained through latitude and longitude coordinates carried in remote sensing images.

[0078] Weather conditions: can be determined by image sharpness evaluation functions (such as Brenner gradient) or dedicated classification models, and are divided into "sunny", "foggy", "rainy" and "snowy".

[0079] Target feature quantification: mainly targets the target itself, for example:

[0080] Target scale (pixel ratio): Calculates the ratio of the target bounding box area to the total image area, and classifies it into "small target", "medium target" and "large target".

[0081] Target types: ships, aircraft, vehicles, etc.

[0082] Target camouflage characteristics quantification: assess the type of target camouflage and whether it is camouflaged at all.

[0083] In some embodiments of the present invention, the quantitative classification of each influencing factor includes:

[0084] For continuous influencing factors, their value range is divided into multiple levels, and a numerical range or descriptive label is set for each level;

[0085] For discrete influencing factors, their different values ​​are directly used as different levels;

[0086] In some embodiments of the present invention, constructing multiple evaluation datasets with different combinations of influencing factor labels includes:

[0087] For each image in the original image database, based on its imaging conditions and content characteristics, the quantitative level of each influencing factor is labeled according to the scene influencing factor system to form a combination label of the influencing factors of the image;

[0088] Before annotation, annotation standards need to be established. A detailed annotation manual should be created based on five major categories of influencing factors and their quantification levels: platform characteristics, load characteristics, environmental characteristics, target characteristics, and target camouflage characteristics. The manual should provide typical example images for each level to reduce annotation subjectivity. At the same time, routine target location (bounding box) and category annotation are still required.

[0089] When annotating, a "human-machine collaboration" approach can be adopted. First, use metadata parsing tools to automatically extract some information from each image, such as obtaining focal length, aperture, and ISO from EXIF ​​data, and associating platform speed information from logs. Then, the annotator displays the image and the parsed information. The annotator needs to complete the following tasks: (1) confirm or correct the level corresponding to the automatically extracted information; (2) for factors that cannot be automatically obtained (such as the degree of target camouflage or occlusion), manually determine and select the level according to the annotation manual; (3) complete the bounding box and category annotation of all target instances in the image.

[0090] Simultaneously, a data cleaning and verification process is required to check the consistency of the annotation results. For example, the annotation results of similar images from the same batch by different annotators are compared, and samples with excessively large differences are eliminated. The logical rationality of the annotation is checked, such as situations where the "lighting conditions" are "low illumination" but the "average image brightness" is very high.

[0091] Images are classified based on the combination of influencing factors labels, and images with the same or similar combination of influencing factors labels are grouped together to form multiple evaluation sub-datasets.

[0092] In some embodiments of the present invention, the systematic evaluation of each target detection model in the initial model library includes:

[0093] Each evaluation subset is input into each object detection model to obtain the model's detection results;

[0094] For each model's detection results on each evaluation dataset, multiple evaluation metrics are calculated, including at least a detection accuracy metric and a computational efficiency metric.

[0095] A model performance matrix is ​​established based on evaluation metrics. The rows of the matrix represent different object detection models, the columns represent different evaluation datasets, and the matrix elements are the comprehensive evaluation scores of the models on the corresponding datasets.

[0096] In some embodiments of the present invention, determining the optimal target detection model corresponding to each combination of influencing factors includes:

[0097] For each evaluation dataset corresponding to a specific combination of influencing factors, the target detection model with the highest comprehensive evaluation score in that column of the model performance matrix is ​​selected as the optimal target detection model for that combination of influencing factors.

[0098] When the combined evaluation scores of multiple models differ within a preset threshold, the model with higher computational efficiency is selected as the optimal target detection model.

[0099] In some embodiments of the present invention, establishing the mapping table between the combination of influencing factors and the optimal target detection model includes:

[0100] Construct a multidimensional lookup table, where the index of the table is the code of the combination of influencing factors, and the table entries are the identification information and performance parameters of the corresponding optimal target detection model;

[0101] The mapping table is stored in the local storage of the edge device or in a cloud server and is provided for querying via an API interface.

[0102] In some embodiments of the present invention, step S4 specifically includes:

[0103] Step S41: Set evaluation indicators and comprehensive scoring function.

[0104] Evaluation metrics should balance accuracy and efficiency. For example, accuracy metrics: (Average precision when IoU threshold is 0.5) (Average mAP across multiple IoU thresholds). Efficiency metrics: Average inference latency (milliseconds) or frames per second on the target edge hardware (e.g., a specific mobile GPU or NPU model). ), model file size ( (Unit: MB)

[0105] Design a comprehensive scoring function F. Since different applications prioritize accuracy and efficiency differently, this function should be configurable. For example, a configurable function would look like this:

[0106]

[0107] ,

[0108] in, This means normalizing the original index values ​​to... interval; Let be the weighting coefficient, satisfying For applications requiring high precision, the height can be adjusted. For applications with extremely high real-time requirements, the weight can be increased. .

[0109] Step S42: Perform batch evaluation. Each model is loaded automatically in sequence. In each evaluation dataset Perform inference and calculate the aforementioned metrics. This process can be parallelized to accelerate it. Record all results to form the original evaluation log.

[0110] Step S43: Calculate the performance matrix. Based on the comprehensive scoring function from step S41, calculate the performance matrix for each model-dataset pair. , fill the performance matrix P.

[0111] Step S44: Determine the optimal model under each condition. For each column j of the performance matrix P (corresponding to condition...) ), find in this column Maximum value: Model index that obtains the maximum value That is, the optimal model under this condition. If there is a tie for first place, then the selection can be based on secondary principles (such as choosing from a tie-and-match model). (The higher one) makes the decision.

[0112] Step S45: Construct and store the mapping table T. The mapping table T can be stored in a key-value database, a JSON file, or a specific binary format. Each record must contain at least:

[0113] Key: Combination of influencing factors The encoding or feature vector.

[0114] Value: Optimal Model Unique identifiers (such as model file names, hash values), and their comprehensive scores and key performance indicators ( , (etc.) for reference.

[0115] To support fast approximate matching during the detection phase, an index structure can be built for the mapping table T, such as based on a KD-Tree or a spherical tree, to perform nearest neighbor search on continuous or multidimensional keys.

[0116] In some embodiments of the present invention, a mapping table update step is also included:

[0117] Regularly collect new evaluation data and newly released object detection models;

[0118] Re-execute the model evaluation and mapping table construction steps to generate an updated mapping table;

[0119] The updated mapping table is pushed to the edge device using an incremental update method.

[0120] In some embodiments of the present invention, when a query of the mapping table does not find a completely matching combination of influencing factors, one of the following strategies is adopted:

[0121] Select the target detection model corresponding to the closest combination of influencing factors;

[0122] From target detection models corresponding to multiple similar combinations of influencing factors, model fusion is performed based on similarity weights;

[0123] The online evaluation module is activated to select multiple candidate models from the model library for rapid evaluation, determine the most suitable model, and update the mapping table.

[0124] During the detection phase, after analyzing and obtaining the vector of influencing factors for the new scene, this vector is used as the key to query the mapping table T. Two scenarios may occur:

[0125] Scenario A: Exact Match. Find the key in T that is exactly the same as the vector of influencing factors in the new scenario, and directly return the corresponding optimal model. This is the ideal situation.

[0126] Scenario B: No exact match. Because real-world scenarios are infinitely diverse, and the constructed evaluation dataset can only cover a limited number of combinations of influencing factors, the influencing factor vector for a new scenario is likely to be a new combination not predefined in T. In this case, one of the following backup strategies needs to be adopted:

[0127] Strategy B1: Nearest Neighbor Matching. Calculate the distance (e.g., Euclidean distance, Hamming distance, or cosine distance, depending on the encoding method) between the vector of influencing factors in the new scene and all keys (vectors of combined influencing factors) in T. Select the model corresponding to the closest key as the approximate optimal model for the current scene. This strategy is suitable for scenarios with continuous scene changes.

[0128] Strategy B2: Similarity-based model fusion. Find the K scene combinations {C1, C2, ..., Ck} and their corresponding models {m1, m2, ..., mk} that are closest to the influencing factor vectors of the new scene. Calculate weights {w1, w2, ..., wk} based on the reciprocal of the distance or the similarity score. Then, use a weighted fusion method (such as weighted fusion of detection result boxes, or adaptive ensemble at the model weight level) to generate a "virtual optimal model" for the current new scene. This strategy allows for a smooth transition and may achieve better performance than a single nearest neighbor model.

[0129] Strategy B3: Trigger a lightweight online evaluation. If the edge device has sufficient computing resources at the current moment, a rapid evaluation process can be initiated. N candidate models are pre-selected from the model library based on the similarity between the influencing factor vector of the new scene and various scene combinations. These models are then quickly run on a subset or downsampled version of the current image. A temporary optimal model is selected based on a simplified evaluation metric (such as a preliminary judgment of inference speed and the number of detection boxes). Simultaneously, the {influencing factor vector of the new scene, selected model} record can be fed back to the cloud for future offline updates to the mapping table T.

[0130] Strategy B4: Default model downgrade. When none of the above strategies apply or resources are extremely limited, the system falls back to a preset, relatively strong default lightweight model (e.g., the model with the highest average score across all datasets when building the mapping table).

[0131] This invention also provides a multi-model target detection system driven by influencing factors in complex dynamic scenes, used to implement the above-mentioned method, such as... Figure 2 As shown, it includes:

[0132] The model library management module is used to maintain a model library containing multiple lightweight object detection models;

[0133] The scenario analysis module is used to analyze complex dynamic scenarios, breaking down the scenario into multiple quantifiable influencing factors and quantifying and classifying them.

[0134] The evaluation dataset construction module is used to construct multiple evaluation datasets with different combinations of influencing factor labels;

[0135] The model evaluation module is used to evaluate the target detection models in the model library using the evaluation dataset;

[0136] The mapping table construction module is used to establish a mapping relationship table between the combination of influencing factors and the optimal target detection model based on the evaluation results;

[0137] The model selection module is used to query the mapping relationship table based on the combination of factors affecting the new task and select the corresponding optimal target detection model.

[0138] The detection execution module is used to perform object detection tasks using the selected object detection model.

[0139] According to one aspect of the present invention, an electronic device is provided, comprising: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and the one or more computer programs are stored in the memory; when the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform a multi-model target detection method driven by influencing factors in a complex dynamic scene as described in any of the above technical solutions.

[0140] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0141] The memory can be an internal storage unit of the terminal device, such as a hard drive or RAM. Alternatively, it can be an external storage device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory can include both internal and external storage units. The memory is used to store the computer program and other programs and data required by the terminal device. It can also be used to temporarily store data that has been output or will be output.

[0142] According to one aspect of the present invention, a computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement a multi-model target detection method driven by influencing factors in a complex dynamic scene, as described in any of the above technical solutions.

[0143] For example, computer-readable storage media can be read-only memory (ROM), random access memory (RAM), read-only optical disc (CD-ROM), magnetic tape, floppy disk, and optical data storage devices. They can be implemented using computer-executable program code, thus allowing them to be stored in a storage device for execution by a computing device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Therefore, this invention is not limited to any particular hardware and software combination.

[0144] In summary, this invention proposes a multi-model target detection method driven by influencing factors in complex dynamic scenarios. It adopts an influencing factor classification method and automatically optimizes the existing influencing factor model library to find the optimal model for the target detection task.

[0145] Furthermore, this invention improves existing target detection methods by employing an automatic optimization strategy for influencing factors, thereby comprehensively enhancing target detection capabilities.

[0146] Furthermore, it should be noted that embodiments of the present invention may take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0147] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0148] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0149] It should also be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0150] Finally, it should be noted that the above description represents a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A multi-model target detection method driven by influencing factors in complex dynamic scenes, characterized in that, Includes the following steps: Step S1: Obtain or train multiple lightweight object detection models to build an initial model library; Step S2: Decompose the complex dynamic scene that affects the target detection performance, identify multiple quantifiable scene influencing factors, and quantify and classify each influencing factor. Step S3: Based on the scene influencing factor system after quantification and grading, the original image database is labeled to construct multiple evaluation datasets with different combinations of influencing factor labels; Step S4: Systematically evaluate each target detection model in the initial model library using the multiple evaluation datasets, determine the optimal target detection model corresponding to each combination of influencing factors, and establish a mapping relationship table between the combination of influencing factors and the optimal target detection model; Step S5: For a new target detection task, analyze its corresponding complex scenario, determine the specific combination of influencing factors under the scenario influencing factor system, query the mapping relationship table to select the corresponding optimal target detection model to execute the detection task.

2. The multi-model target detection method driven by influencing factors in complex dynamic scenes according to claim 1, characterized in that, The influencing factors of the scenario include at least three of the following: platform characteristics, load characteristics, environmental characteristics, target characteristics, and target camouflage characteristics.

3. The multi-model target detection method driven by influencing factors in complex dynamic scenes according to claim 2, characterized in that, The platform characteristics include at least one of the following: type of platform, track height, attitude change, and vibration amplitude; The load characteristics include at least one of the following: sensor type, resolution, focal length, field of view, and imaging mode; The environmental characteristics include at least one of the following: country / region, weather conditions, sea breeze, and sea state; The target characteristics include at least one of the following: target size, target attitude, target category, and operating condition; The target camouflage characteristics include at least one of optical camouflage and smoke concealment.

4. The multi-model target detection method driven by influencing factors in complex dynamic scenes according to claim 1, characterized in that, The quantitative classification of each influencing factor includes: For continuous influencing factors, their value range is divided into multiple levels, and a numerical range or descriptive label is set for each level; For discrete influencing factors, their different values ​​are directly used as different levels.

5. The multi-model target detection method driven by influencing factors in complex dynamic scenes according to claim 1, characterized in that, The construction of multiple evaluation datasets with different combinations of influencing factors and labels includes: For each image in the original image database, based on its imaging conditions and content characteristics, the quantitative level of each influencing factor is labeled according to the scene influencing factor system to form a combination label of the influencing factors of the image; Images are classified based on the combination of influencing factors labels, and images with the same or similar combination of influencing factors labels are grouped together to form multiple evaluation sub-datasets.

6. The multi-model target detection method driven by influencing factors in complex dynamic scenes according to claim 1, characterized in that, The systematic evaluation of each target detection model in the initial model library includes: Each evaluation subset is input into each object detection model to obtain the model's detection results; For each model's detection results on each evaluation dataset, multiple evaluation metrics are calculated, including at least a detection accuracy metric and a computational efficiency metric. A model performance matrix is ​​established based on evaluation metrics. The rows of the matrix represent different object detection models, the columns represent different evaluation datasets, and the matrix elements are the comprehensive evaluation scores of the models on the corresponding datasets.

7. The multi-model target detection method driven by influencing factors in complex dynamic scenes according to claim 6, characterized in that, The determination of the optimal target detection model for each combination of influencing factors includes: For each evaluation dataset corresponding to a specific combination of influencing factors, the target detection model with the highest comprehensive evaluation score in that column of the model performance matrix is ​​selected as the optimal target detection model for that combination of influencing factors. When the combined evaluation scores of multiple models differ within a preset threshold, the model with higher computational efficiency is selected as the optimal object detection model.

8. The multi-model target detection method driven by influencing factors in complex dynamic scenes according to claim 1, characterized in that, The mapping table between the combination of influencing factors and the optimal target detection model includes: Construct a multidimensional lookup table, where the index of the table is the code of the combination of influencing factors, and the table entries are the identification information and performance parameters of the corresponding optimal target detection model; The mapping table is stored in the local storage of the edge device or in a cloud server and is provided for querying via an API interface.

9. The multi-model target detection method driven by influencing factors in complex dynamic scenes according to claim 1, characterized in that, It also includes the mapping table update step: Regularly collect new evaluation data and newly released object detection models; Re-execute the model evaluation and mapping table construction steps to generate an updated mapping table; The updated mapping table is pushed to the edge device using an incremental update method.

10. The multi-model target detection method driven by influencing factors in complex dynamic scenes according to claim 1, characterized in that, When the query of the mapping table does not find a completely matching combination of influencing factors, one of the following strategies is adopted: Select the target detection model corresponding to the closest combination of influencing factors; From target detection models corresponding to multiple similar combinations of influencing factors, model fusion is performed based on similarity weights; The online evaluation module is activated to select multiple candidate models from the model library for rapid evaluation, determine the most suitable model, and update the mapping table.

Images