An adaptive image target detection algorithm selection method based on comprehensive evaluation

Abstract

The application relates to an adaptive image target detection algorithm selection method based on comprehensive evaluation, and belongs to the technical field of target detection. According to the requirements and characteristics of video structured processing, an evaluation model is established according to the type and multi-attribute of a target detection algorithm, an adaptive algorithm selection method based on actual scene and algorithm attribute matching is proposed, the accuracy and precision of target detection can be improved, and the method better serves subsequent related applications and deep analysis and processing about the target.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology, specifically relating to an adaptive image target detection algorithm selection method based on comprehensive evaluation. Background Technology

[0002] Object detection is a computer vision task designed to identify specific objects in an image or video and determine their location. This requires not only identifying the object's category but also drawing its bounding box. It plays an important role in autonomous driving, security monitoring, and medical image analysis.

[0003] Existing object detection methods can be broadly categorized into traditional image processing-based methods and deep learning-based methods. Each category includes several different approaches. For example, traditional image processing-based object detection methods include frame differencing and CascadeClassifier, while deep learning-based methods include the YOLO series and Fast R-CNN. Each method has its own usage conditions and applicable scope, such as requirements for the environment and operating components, resource consumption, and running speed. Furthermore, no single method can guarantee sufficient accuracy in a specific scenario; false positives and false negatives are possible.

[0004] In general, object detection method selection is static. This means that a single object detection method is selected and evaluated based on the operating scenario and performance requirements, and a particular algorithm is chosen to run in that scenario or environment. Once the detection algorithm is determined, it is not modified or adjusted. This approach has drawbacks: it cannot be dynamically adjusted, a single algorithm may not meet performance requirements, and computational resources are wasted.

[0005] The problems that need to be solved are how to adaptively select the appropriate object detection algorithm according to a certain scenario, and how to improve the accuracy of these models or algorithms in a certain scenario. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] The technical problem to be solved by this invention is to provide a method for selecting an image target detection algorithm to improve the accuracy and precision of target detection, and better serve subsequent target-related applications and in-depth analysis and processing.

[0008] (II) Technical Solution

[0009] To address the aforementioned technical problems, this invention provides an adaptive image target detection algorithm selection method based on comprehensive evaluation, comprising the following steps:

[0010] Step 1: Describe the target detection algorithm using multiple attributes and establish multi-attribute evaluation rules for target detection;

[0011] Video structuring is the process of converting unstructured video streams into structured data, which refers to the target information contained in the video. The processing methods for target detection are divided into traditional image processing methods and deep learning methods, where "0" represents traditional image processing methods and "1" represents deep learning methods. Representative specific methods are selected for each category, with no more than 10 categories, and the specific codes are represented by single-digit numbers from 0 to 9.

[0012] Each type of method has its own attributes:

[0013] 1) Coding languages: divided into two categories: C and Python;

[0014] 2) Component runtime environment: Includes the relevant component name and version number, where the version number is a range, divided into three cases: higher than a certain version and lower than a certain version, higher than a certain version, and lower than a certain version; View the current version number of the component, and count the number of supported versions based on the version number range;

[0015] (1) Total number of components required: The more components required, the lower the score; with 100 components as the upper limit, [0,10] gets 1 point, (10,20] gets 0.9 points, (20,30] gets 0.8 points, and so on, (90,100] gets 0.1 points, and more than 100 gets 0 points;

[0016] (2) Regular components: quantity and version range; the more components, the lower the score; the wider the version range, the higher the score; the number of components is limited to 10, <=1 component gets 1 point, (1,2] components get 0.9 points, and so on, more than 10 components get 0 points; the supported version number, more than 10 components get 1 point, less than 10 components, and so on, supporting 9 components gets 0.9 points;

[0017] (3) Special components: quantity and version range; the more the quantity, the lower the score; the larger the version range, the higher the score; the number of components is capped at 90, <=10 components get 1 point, (10,20] components get 0.8 points, and so on, more than 60 components get 0 points; the number of supported version numbers, more than 10 components get 1 point, less than 10 components, and so on, supporting 9 components gets 0.9 points;

[0018] 3) The running speed and resource consumption of a computer with a specific configuration;

[0019] (1) The faster the processing speed of a single image, the higher the score;

[0020] In a GPU-enabled environment, a time less than 0.001 seconds earns 1 point, [0.001, 0.01] earns 0.8 points, (0.01, 0.02] earns 0.6 points, (0.02, 0.04] earns 0.4 points, (0.04, 0.08] earns 0.2 points, and a time greater than 0.08 seconds earns 0 points.

[0021] In an environment without a GPU, a time of less than 0.01 seconds earns 1 point, [0.01, 0.05] earns 0.8 points, (0.05, 0.1] earns 0.6 points, (0.1, 0.2] earns 0.4 points, (0.2, 0.4] earns 0.2 points, and a time of more than 0.4 seconds earns 0 points.

[0022] (2) Resource usage includes CPU and memory usage without GPU and CPU, memory and GPU usage with GPU. GPU usage is further divided into video memory usage and computing power usage. The lower the video memory usage, the higher the score; the lower the computing power usage, the higher the score.

[0023] 0.2G of video memory usage earns 1 point, 0.5G earns 0.8 points, 1G earns 0.6 points, 1.5G earns 0.4 points, 2G earns 0.2 points, and more than 2G earns 0 points.

[0024] 4) Performance on a specific dataset, including the dataset name, set of metrics, and corresponding values, with a weighted comprehensive evaluation of the performance metric values;

[0025] MAP>90: 1 point;>85 and≤90: 0.8 points;>80 and≤85: 0.6 points;>75 and≤80: 0.4 points;>70 and≤75: 0.2 points; below 70: 0 points.

[0026] Establish evaluation rules for the evaluation factors and obtain the evaluation results for each method;

[0027] Step 2: Selection of single target detection algorithm;

[0028] Based on step 1, design corresponding strategies according to requirements, and select an applicable single model or algorithm according to environmental and performance requirements.

[0029] (III) Beneficial Effects

[0030] This invention addresses the needs and characteristics of video structured processing. Based on the type of target detection algorithm and the establishment of an evaluation model with multiple attributes, it proposes an adaptive algorithm selection method that matches the actual scene with the algorithm attributes. This method can improve the accuracy and precision of target detection and better serve subsequent target-related applications and in-depth analysis and processing. Detailed Implementation

[0031] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to examples.

[0032] There are many object detection methods, and the choice of which algorithm(s) to use in a real-world environment affects the speed and accuracy of the results. It is necessary to match the applicable scenarios of object detection with the actual scenarios. Therefore, this invention provides an adaptive image object detection algorithm selection method based on comprehensive evaluation. It describes object detection methods from multiple perspectives and establishes evaluation methods and metrics. Based on this, one or more algorithms are matched according to the actual scenario. The method specifically includes the following steps:

[0033] Step 1: Describe the object detection algorithm using multiple attributes and establish multi-attribute evaluation rules for object detection.

[0034] Video structuring transforms unstructured video streams into structured data, where structured data refers to the target information contained within the video. Object detection methods are categorized into traditional image processing methods and deep learning-based methods, where "0" represents traditional image processing methods and "1" represents deep learning-based methods. Representative methods are selected for each category, with no more than 10 categories, and their specific codes are represented by single-digit numbers from 0 to 9.

[0035] Each type of method has its own attributes:

[0036] 1. Coding languages: Divided into two categories: C and Python;

[0037] 2. Component runtime environment: This includes the relevant component name and version number. The version number is a range, categorized into three cases: higher than a certain version and lower than a certain version, higher than a certain version, and lower than a certain version. Generally, you can check the current version number information for the component and, based on the version number range, count the number of supported versions.

[0038] (1) Total number of components required: The more components required, the lower the score. With 100 components as the upper limit, [0,10] gets 1 point, (10,20] gets 0.9 points, (20,30] gets 0.8 points, and so on, (90,100] gets 0.1 points, and more than 100 gets 0 points;

[0039] (2) Regular Components: Quantity and version range. The more components, the lower the score; the wider the version range, the higher the score. The maximum number of components is 10. 1 component <= 1 gets 1 point, (1,2] gets 0.9 points, and so on. More than 10 components get 0 points. Supported version numbers: more than 10 get 1 point, less than 10 get a decreasing score, such as supporting 9 get 0.9 points.

[0040] (4) Special components: quantity and version range. The more components, the lower the score; the wider the version range, the higher the score. The number of components is capped at 90. 1 point is awarded for <=10 components, 0.8 points for (10,20] components, and so on. 0 points are awarded for more than 60 components. Supported version numbers are awarded as follows: 1 point is awarded for more than 10 versions, and the score decreases sequentially for less than 10 versions. For example, 0.9 points are awarded for supporting 9 versions.

[0041] 3. Running speed and resource consumption under a specific computer configuration.

[0042] (1) The processing time for a single image, in frames per second. The faster the speed, the higher the score.

[0043] In a GPU-enabled environment, a time less than 0.001 seconds earns 1 point, [0.001, 0.01] earns 0.8 points, (0.01, 0.02] earns 0.6 points, (0.02, 0.04] earns 0.4 points, (0.04, 0.08] earns 0.2 points, and a time greater than 0.08 seconds earns 0 points.

[0044] In an environment without a GPU, a time of less than 0.01 seconds earns 1 point, [0.01, 0.05] earns 0.8 points, (0.05, 0.1] earns 0.6 points, (0.1, 0.2] earns 0.4 points, (0.2, 0.4] earns 0.2 points, and a time of more than 0.4 seconds earns 0 points.

[0045] (2) Resource usage includes CPU and memory usage without GPU and CPU, memory and GPU usage with GPU. GPU usage is further divided into video memory usage and computing power usage. The lower the video memory usage, the higher the score; the lower the computing power usage, the higher the score.

[0046] 0.2G of video memory usage earns 1 point, 0.5G earns 0.8 points, 1G earns 0.6 points, 1.5G earns 0.4 points, 2G earns 0.2 points, and more than 2G earns 0 points.

[0047] 4. Performance on a specific dataset, including the dataset name, set of metrics, and corresponding values. A weighted comprehensive evaluation of the performance metric values.

[0048] MAP>90 gets 1 point,>85 gets 0.8 points,>80 gets 0.6 points,>75 gets 0.4 points,>70 gets 0.2 points, and below70 gets 0 points.

[0049] Establish evaluation rules for the above evaluation factors and obtain the evaluation results for each method.

[0050] Step 2: Selection of Single Object Detection Algorithm

[0051] Based on step 1, design corresponding strategies according to requirements, and select a suitable single model or algorithm according to environmental and performance requirements. The steps are as follows:

[0052] Step 1: Select a set of algorithms that meet the requirements based on language constraints;

[0053] Step 2: Filter the algorithm set based on the matching degree between the existing environment and the required environment;

[0054] Step 3: Filter the algorithm set based on the degree of matching in terms of speed and resource consumption;

[0055] Step 4: Filter the algorithm set according to the degree of performance matching;

[0056] Step 5: Select a set of algorithms that meet the requirements from the remaining set based on the score constraints;

[0057] Step 6: Select the algorithm with the highest score.

[0058] If the algorithm set is empty after screening, then an objective optimization function is established to select the algorithm with the highest score. The method is as follows:

[0059] Assume the following factors: score difference, difference in the number of commonly used components, difference in the number of special components, proportional coefficient of the version range of commonly used components, proportional coefficient of the template range of special components, proportional coefficient of the difference in indicators, proportional coefficient of speed difference, and proportional coefficient of the difference in the common use of resources; calculate the weighted sum and select the algorithm with the smallest value.

[0060] Steps 1 and 2 form the first solution.

[0061] Step 3: Adaptively select multiple object detection algorithms

[0062] Based on steps 1 and 2, corresponding strategies are designed according to requirements, and multiple models or algorithms are adaptively selected for detection according to environmental and performance requirements to improve target detection accuracy. A result fusion strategy is also designed.

[0063] If high accuracy is required, the steps are as follows:

[0064] Step 31: Based on available resources, select candidate algorithms and calculate the number of algorithms that can run simultaneously.

[0065] Step 32: Set a false negative rate threshold, and then select algorithms from the candidate algorithms that are lower than the false negative rate threshold;

[0066] Step 33: Select the algorithm with the highest accuracy, and reduce the number of algorithms that can be run simultaneously by 1;

[0067] Step 34: If the number of runnable algorithms is not zero, continue to select the algorithm with the next lowest accuracy; check for duplicates. Broken Until the number of runnable algorithms reaches 0;

[0068] Step 35: Perform fusion processing on the detection results:

[0069] Step 351: Set a threshold and select the results that meet the criteria;

[0070] Step 352: Filter out duplicate results and merge IOUs;

[0071] Step 353: Sort according to confidence level;

[0072] Step 354: Select a target and iterate through all rectangles of the same target type within the domain;

[0073] Step 355: Calculate the IOU value between each rectangle and the reference rectangle;

[0074] Step 356: If the IOU value is greater than the set threshold, it is considered a duplicate and merged.

[0075] If a low false negative rate is required, the steps are as follows:

[0076] Step 31': Based on available resources, select candidate algorithms and calculate the number of algorithms that can run simultaneously.

[0077] Step 32': Set an accuracy threshold and filter out algorithms with higher accuracy from the candidate algorithms again;

[0078] Step 33': Select the algorithm with the lowest false negative rate, and reduce the number of algorithms that can be run simultaneously by 1;

[0079] Step 34': If the number of runnable algorithms is not 0, continue to select the algorithm with the second lowest false negative rate; repeat the judgment until the number of runnable algorithms is 0.

[0080] Step 35': Perform fusion processing on the detection results:

[0081] Step 351': Set a threshold and select the results that meet the criteria;

[0082] Step 352': Filter out duplicate results and merge IOUs;

[0083] Steps 1 and 3 form the second solution.

[0084] Step 4: Group the target detection algorithms and adaptively select the grouping algorithm.

[0085] Grouping strategies are designed for relevant object detection algorithms. Based on environmental and performance requirements, a grouping selection strategy and a result fusion strategy are designed. Complementary algorithms are selected based on the complementarity of results on the dataset. Building upon step 1, this step performs complementarity and correlation analysis on the object detection algorithms and selects different grouping algorithms according to different performance requirements.

[0086] Assume the sample set contains a collection of images. ,in Let be the number of samples in the set. For simplicity and to avoid affecting the results, assume that each image sample contains one object, and use three object detection algorithms, A, B, and C, as examples. We will then use each of the three object detection algorithms to detect the object on the sample set. Assume that the set of correctly detected objects is . The set of common errors or missed detections is Of the remaining test results, This indicates that algorithm A detected correctly, but algorithm B detected incorrectly or missed detections. This indicates that Algorithm A detects correctly and Algorithm B also detects correctly, and so on. The results of the three algorithms, except... , In addition, a total of 9 sets were generated. , , , , , , , , The first three sets are considered as substitutable sets, and the last six sets are considered as complementary sets.

[0087] For the first three sets, count the number of samples in each set. > This indicates that for algorithm A, B is a stronger substitute than C, and so on.

[0088] For the last 6 sets, count the number of samples in each set. > This indicates that for algorithm B, the complementarity of A is stronger than that of C, and so on.

[0089] Establish a two-dimensional matrix of substitutability and complementarity, and assign corresponding values ​​according to the strength of substitutability and complementarity.

[0090] Regarding substitutability, for algorithm A, ,when hour, , ;when hour, , ;when hour, , ;when hour, , ;when hour, , .

[0091] Regarding complementarity, for algorithm A, ,when hour, , . To be modified as needed.

[0092] Assuming the set of correctly detected results is , , , The total number of errors detected was , , , .

[0093] To achieve a high accuracy rate, the steps are as follows:

[0094] Step 41: Based on available resources, filter out combinations that meet the criteria;

[0095] Step 42: Select the combination with the highest union of correct results;

[0096] Step 43: Perform the test;

[0097] Step 44: Filter out the set of results with a confidence level higher than the threshold;

[0098] Step 45: Merge IOUs.

[0099] If a low false negative rate is required, the steps are as follows:

[0100] Step 41': Based on available resources, filter out combinations that meet the criteria;

[0101] Step 42': Select the combination with the fewest result union errors;

[0102] Step 43': Perform the test;

[0103] Step 44': Filter out the set of results with a confidence level higher than the threshold;

[0104] Step 45': Merge the IOUs.

[0105] Steps 1 and 4 form the third solution.

[0106] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An adaptive image target detection algorithm selection method based on comprehensive evaluation, characterized in that, Includes the following steps: Step 1: Describe the target detection algorithm using multiple attributes and establish multi-attribute evaluation rules for target detection; Video structuring is the process of converting unstructured video streams into structured data. Structured data refers to the target information contained in the video. The processing methods for target detection are divided into methods based on traditional image processing and methods based on deep learning. "0" represents traditional image processing methods and "1" represents deep learning methods. Representative specific methods are selected for each category, with no more than 10 categories. The specific codes are represented by single-digit numbers from 0 to 9. Each type of method has its own attributes: 1) Coding languages: divided into two categories: C and Python; 2) Component runtime environment: Includes the relevant component name and version number, where the version number is a range, divided into three cases: higher than a certain version and lower than a certain version, higher than a certain version, and lower than a certain version; View the current version number of the component, and count the number of supported versions based on the version number range; (1) Total number of components required: The more components required, the lower the score; with 100 components as the upper limit, [0,10] gets 1 point, (10,20] gets 0.9 points, (20,30] gets 0.8 points, and so on, (90,100] gets 0.1 points, and more than 100 gets 0 points; (2) Regular components: quantity and version range; the more components, the lower the score; the wider the version range, the higher the score; the number of components is limited to 10, <=1 component gets 1 point, (1,2] components get 0.9 points, and so on, more than 10 components get 0 points; the supported version number, more than 10 components get 1 point, less than 10 components, and so on, supporting 9 components gets 0.9 points; (3) Special components: quantity and version range; the more the quantity, the lower the score; the larger the version range, the higher the score; the number of components is capped at 90, <=10 components get 1 point, (10,20] components get 0.8 points, and so on, more than 60 components get 0 points; the number of supported version numbers, more than 10 components get 1 point, less than 10 components, and so on, supporting 9 components gets 0.9 points; 3) The running speed and resource consumption of a computer with a specific configuration; (1) The faster the processing speed of a single image, the higher the score; In a GPU-enabled environment, a time less than 0.001 seconds earns 1 point, [0.001, 0.01] earns 0.8 points, (0.01, 0.02] earns 0.6 points, (0.02, 0.04] earns 0.4 points, (0.04, 0.08] earns 0.2 points, and a time greater than 0.08 seconds earns 0 points. In an environment without a GPU, a time of less than 0.01 seconds earns 1 point, [0.01, 0.05] earns 0.8 points, (0.05, 0.1] earns 0.6 points, (0.1, 0.2] earns 0.4 points, (0.2, 0.4] earns 0.2 points, and a time of more than 0.4 seconds earns 0 points. (2) Resource usage includes CPU and memory usage without GPU and CPU, memory and GPU usage with GPU. GPU usage is further divided into video memory usage and computing power usage. The lower the video memory usage, the higher the score; the lower the computing power usage, the higher the score. 0.2G of video memory usage earns 1 point, 0.5G earns 0.8 points, 1G earns 0.6 points, 1.5G earns 0.4 points, 2G earns 0.2 points, and more than 2G earns 0 points. 4) Performance on a specific dataset, including the dataset name, set of metrics, and corresponding values, with a weighted comprehensive evaluation of the performance metrics values; MAP>90: 1 point;>85 and≤90: 0.8 points;>80 and≤85: 0.6 points;>75 and≤80: 0.4 points;>70 and≤75: 0.2 points; below 70: 0 points. Establish evaluation rules for the evaluation factors and obtain the evaluation results for each method; Step 2: Selection of single target detection algorithm; Based on step 1, design corresponding strategies according to requirements, and select an applicable single model or algorithm according to environmental and performance requirements.

2. The method as described in claim 1, characterized in that, Step 2 is as follows: Step 1: Select a set of algorithms that meet the requirements based on language constraints; Step 2: Filter the algorithm set based on the matching degree between the existing environment and the required environment; Step 3: Filter the algorithm set based on the degree of matching in terms of speed and resource consumption; Step 4: Filter the algorithm set according to the degree of performance matching; Step 5: Select a set of algorithms that meet the requirements from the remaining set based on the score constraints; Step 6: Select the algorithm with the highest score; If the algorithm set is empty after screening, then establish an objective optimization function to select the algorithm with the highest score. The method is as follows: Assume differences in scores, differences in the number of commonly used components, differences in the number of special components, proportional coefficients for the version range of commonly used components, proportional coefficients for the template range of special components, proportional coefficients for differences in indicators, proportional coefficients for differences in speed, and proportional coefficients for differences in common resources; The algorithm uses weighted summation to select the minimum value.

3. An adaptive image target detection algorithm selection method based on comprehensive evaluation, characterized in that, Includes the following steps: Step 1: Describe the target detection algorithm using multiple attributes and establish multi-attribute evaluation rules for target detection; Video structuring is the process of converting unstructured video streams into structured data. Structured data refers to the target information contained in the video. The processing methods for target detection are divided into methods based on traditional image processing and methods based on deep learning. "0" represents traditional image processing methods and "1" represents deep learning methods. Representative specific methods are selected for each category, with no more than 10 categories. The specific codes are represented by single-digit numbers from 0 to 9. Each type of method has its own attributes: 1) Coding languages: divided into two categories: C and Python; 2) Component runtime environment: Includes the relevant component name and version number, where the version number is a range, divided into three cases: higher than a certain version and lower than a certain version, higher than a certain version, and lower than a certain version; View the current version number of the component, and count the number of supported versions based on the version number range; (1) Total number of components required: The more components required, the lower the score; with 100 components as the upper limit, [0,10] gets 1 point, (10,20] gets 0.9 points, (20,30] gets 0.8 points, and so on, (90,100] gets 0.1 points, and more than 100 gets 0 points; (2) Regular components: quantity and version range; the more components, the lower the score; the wider the version range, the higher the score; the number of components is limited to 10, <=1 component gets 1 point, (1,2] components get 0.9 points, and so on, more than 10 components get 0 points; the supported version number, more than 10 components get 1 point, less than 10 components, and so on, supporting 9 components gets 0.9 points; (3) Special components: quantity and version range; the more the quantity, the lower the score; the larger the version range, the higher the score; the number of components is capped at 90, <=10 components get 1 point, (10,20] components get 0.8 points, and so on, more than 60 components get 0 points; the number of supported version numbers, more than 10 components get 1 point, less than 10 components, and so on, supporting 9 components gets 0.9 points; 3) The running speed and resource consumption of a computer with a specific configuration; (1) The faster the processing speed of a single image, the higher the score; In a GPU-enabled environment, a time less than 0.001 seconds earns 1 point, [0.001, 0.01] earns 0.8 points, (0.01, 0.02] earns 0.6 points, (0.02, 0.04] earns 0.4 points, (0.04, 0.08] earns 0.2 points, and a time greater than 0.08 seconds earns 0 points. In an environment without a GPU, a time of less than 0.01 seconds earns 1 point, [0.01, 0.05] earns 0.8 points, (0.05, 0.1] earns 0.6 points, (0.1, 0.2] earns 0.4 points, (0.2, 0.4] earns 0.2 points, and a time of more than 0.4 seconds earns 0 points. (2) Resource usage includes CPU and memory usage without GPU and CPU, memory and GPU usage with GPU. GPU usage is further divided into video memory usage and computing power usage. The lower the video memory usage, the higher the score; the lower the computing power usage, the higher the score. 0.2G of video memory usage earns 1 point, 0.5G earns 0.8 points, 1G earns 0.6 points, 1.5G earns 0.4 points, 2G earns 0.2 points, and more than 2G earns 0 points. 4) Performance on a specific dataset, including the dataset name, set of metrics, and corresponding values, with a weighted comprehensive evaluation of the performance metrics values; MAP>90: 1 point;>85 and≤90: 0.8 points;>80 and≤85: 0.6 points;>75 and≤80: 0.4 points;>70 and≤75: 0.2 points; below 70: 0 points. Establish evaluation rules for the evaluation factors and obtain the evaluation results for each method; Step 2: Adaptively select multiple target detection algorithms; Based on the requirements, design corresponding strategies, adaptively select multiple models or algorithms for detection according to environmental and performance requirements, and design result fusion strategies.

4. The method as described in claim 3, characterized in that, If high accuracy is required, step 2 is as follows: Step 31: Based on available resources, select candidate algorithms and calculate the number of algorithms that can run simultaneously; Step 32: Set a false negative rate threshold, and then select algorithms from the candidate algorithms that are lower than the false negative rate threshold; Step 33: Select the algorithm with the highest accuracy, and reduce the number of algorithms that can be run simultaneously by 1; Step 34: If the number of runnable algorithms is not zero, continue to select the algorithm with the next lowest accuracy; check for duplicates. Broken This continues until the number of runnable algorithms reaches 0. Step 35: Perform fusion processing on the detection results: Step 351: Set a threshold and select the results that meet the criteria; Step 352: Filter out duplicate results and merge IOUs; Step 353: Sort according to confidence level; Step 354: Select a target and iterate through all rectangles of the same target type within the domain; Step 355: Calculate the IOU value between each rectangle and the reference rectangle; Step 356: If the IOU value is greater than the set threshold, it is considered a duplicate and merged.

5. The method as described in claim 3, characterized in that, If a low false negative rate is required, step 2 is as follows: Step 31': Based on available resources, select candidate algorithms and calculate the number of algorithms that can run simultaneously; Step 32': Set an accuracy threshold and filter out algorithms with higher accuracy from the candidate algorithms again; Step 33': Select the algorithm with the lowest false negative rate, and reduce the number of algorithms that can be run simultaneously by 1; Step 34': If the number of runnable algorithms is not 0, continue to select the algorithm with the second lowest false negative rate; repeat the judgment until the number of runnable algorithms is 0; Step 35': Perform fusion processing on the detection results: Step 351': Set a threshold and select the results that meet the criteria; Step 352': Filter out duplicate results and merge IOUs.

6. A method for selecting an adaptive image target detection algorithm based on comprehensive evaluation, characterized in that, Includes the following steps: Step 1: Describe the target detection algorithm using multiple attributes and establish multi-attribute evaluation rules for target detection; Video structuring is the process of converting unstructured video streams into structured data. Structured data refers to the target information contained in the video. The processing methods for target detection are divided into methods based on traditional image processing and methods based on deep learning. "0" represents traditional image processing methods and "1" represents deep learning methods. Representative specific methods are selected for each category, with no more than 10 categories. The specific codes are represented by single-digit numbers from 0 to 9. Each type of method has its own attributes: 1) Coding languages: divided into two categories: C and Python; 2) Component runtime environment: Includes the relevant component name and version number, where the version number is a range, divided into three cases: higher than a certain version and lower than a certain version, higher than a certain version, and lower than a certain version; View the current version number of the component, and count the number of supported versions based on the version number range; (1) Total number of components required: The more components required, the lower the score; with 100 components as the upper limit, [0,10] gets 1 point, (10,20] gets 0.9 points, (20,30] gets 0.8 points, and so on, (90,100] gets 0.1 points, and more than 100 gets 0 points; (2) Regular components: quantity and version range; the more components, the lower the score; the wider the version range, the higher the score; the number of components is limited to 10, <=1 component gets 1 point, (1,2] components get 0.9 points, and so on, more than 10 components get 0 points; the supported version number, more than 10 components get 1 point, less than 10 components, and so on, supporting 9 components gets 0.9 points; (3) Special components: quantity and version range; the more the quantity, the lower the score; the larger the version range, the higher the score; the number of components is capped at 90, <=10 components get 1 point, (10,20] components get 0.8 points, and so on, more than 60 components get 0 points; the number of supported version numbers, more than 10 components get 1 point, less than 10 components, and so on, supporting 9 components gets 0.9 points; 3) The running speed and resource consumption of a computer with a specific configuration; (1) The faster the processing speed of a single image, the higher the score; In a GPU-enabled environment, a time less than 0.001 seconds earns 1 point, [0.001, 0.01] earns 0.8 points, (0.01, 0.02] earns 0.6 points, (0.02, 0.04] earns 0.4 points, (0.04, 0.08] earns 0.2 points, and a time greater than 0.08 seconds earns 0 points. In an environment without a GPU, a time of less than 0.01 seconds earns 1 point, [0.01, 0.05] earns 0.8 points, (0.05, 0.1] earns 0.6 points, (0.1, 0.2] earns 0.4 points, (0.2, 0.4] earns 0.2 points, and a time of more than 0.4 seconds earns 0 points. (2) Resource usage includes CPU and memory usage without GPU and CPU, memory and GPU usage with GPU. GPU usage is further divided into video memory usage and computing power usage. The lower the video memory usage, the higher the score; the lower the computing power usage, the higher the score. 0.2G of video memory usage earns 1 point, 0.5G earns 0.8 points, 1G earns 0.6 points, 1.5G earns 0.4 points, 2G earns 0.2 points, and more than 2G earns 0 points. 4) Performance on a specific dataset, including the dataset name, set of metrics, and corresponding values, with a weighted comprehensive evaluation of the performance metric values; MAP>90: 1 point;>85 and≤90: 0.8 points;>80 and≤85: 0.6 points;>75 and≤80: 0.4 points;>70 and≤75: 0.2 points; below 70: 0 points. Establish evaluation rules for the evaluation factors and obtain the evaluation results for each method; Step 2: Group the target detection algorithms and adaptively select the grouping algorithm; Grouping strategies are designed for relevant target detection algorithms. Based on environmental and performance requirements, a grouping selection strategy is designed, and a result fusion strategy is designed. Based on the complementarity of results on the dataset, complementary algorithms are selected.

7. The method as described in claim 6, characterized in that, Step 2: Based on Step 1, perform complementarity and correlation analysis on the target detection algorithms, and select different grouping algorithms according to different performance requirements.