WO2024022450A1

WO2024022450A1 - Scene adaptability improvement method and apparatus for object detection, and object detection system

Info

Publication number: WO2024022450A1
Application number: PCT/CN2023/109613
Authority: WO
Inventors: 赵鑫
Original assignee: 杭州海康威视数字技术股份有限公司
Priority date: 2022-07-27
Filing date: 2023-07-27
Publication date: 2024-02-01
Also published as: CN117523466A

Abstract

Disclosed in the present application are a scene adaptability improvement method and apparatus for object detection, and an object detection system. The method comprises: obtaining a detection result of current object data; performing object feature extraction on the current object data; determining a degree of matching between each object feature and each feature in a feature library; comparing each degree of matching with a detection threshold bound with a matched feature, in the feature library, used for determining the degree of matching; and determining processing of the detection result according to the comparison result, wherein the feature library is a feature set of object data calibrated according to the detection result which indicates a false alarm or a missing alarm; and the detection threshold bound with each matched feature is adjusted along with a degree of matching between the matched feature and the object feature of the object data within a preset range. The present application facilitates the reduction of missing alarms and false alarms, and has high scene adaptability.

Description

Method, device, and target detection system for improving scene adaptability of target detection

This application claims the priority of the Chinese patent application submitted to the China Patent Office on July 27, 2022, with the application number 202210889349. The contents are incorporated into this application by reference.

Technical field

The present invention relates to the field of target detection in security, and in particular, to a method, device, and target detection system for improving scene adaptability of target detection.

Background technique

With the development of image-based target detection technology, image-based target detection technology is used in more and more application scenarios, such as security, industrial environment monitoring, etc. For example, in practical applications, the monitored status is usually reflected based on the detection results of the target data. When the detection results meet the preset conditions, such as when the category and confidence level of the detected target meet the preset conditions, then Trigger alarms and other prompts.

In order to improve the accuracy and reliability of alarm triggering and avoid missed and false alarms, it has become an urgent problem to solve the problem of better adaptability of target detection to application scenarios.

Contents of the invention

The present invention provides a method, a device, and a target detection system for improving the scene adaptability of target detection to improve the adaptability of target detection to application scenarios and reduce false negatives and/or false positives.

A first aspect of the present invention provides a method for improving scene adaptability of target detection, which method includes:

Get the detection results of the current target data;

Extract target features from the current target data;

Determine the matching degree of each target feature with each feature in the feature library;

Compare each matching degree with the detection threshold bound to the matched feature in the feature library used to determine the matching degree;

Determine the processing of the detection results based on the comparison results;

in,

The feature library is a feature set of target data calibrated according to whether the detection result is a false positive or a false negative;

The detection threshold bound to each matched feature is adjusted according to the matching degree between the matched feature and the target feature of the target data within the preset range.

A second aspect of the present invention provides a device for improving scene adaptability of target detection. The device includes:

The target detection module is used to obtain the detection results of the current target data and extract target features from the current target data;

A feature matching module, used to match each target feature with each feature in the feature library, where the feature library is a feature set of target data calibrated according to whether the detection result is a false positive or a false negative;

A comparison module, used to compare the detection results with the detection threshold bound to the matching feature in the feature library;

A determination module, used to determine the processing of the detection result according to the comparison result;

A threshold adjustment module is configured to adjust the detection threshold bound to each matched feature according to the matching degree between the matched feature and the target feature of the target data within the preset range.

A third aspect of the present invention provides a target detection system, including the device for improving scene adaptability of target detection.

This application provides a method for improving the scene adaptability of target detection. The detection result is determined by comparing the matching degree between the target feature and each feature in a specific feature library and the detection threshold bound to each feature in the feature library. This allows the feature data in the feature database to be fully utilized, and the granularity of the detection threshold is more delicate. Moreover, the detection threshold can be adjusted adaptively to the target data, which is conducive to improving the adaptability of the detection threshold to the scene, thereby improving the target detection scene. Adaptability helps reduce false negatives and false positives.

Description of drawings

The drawings described here are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application.

Figure 1 is a flow chart of a method for improving scene adaptability of target detection;

Figure 2 is another schematic flowchart of a method for improving scene adaptability of target detection according to an embodiment of the present application;

Figure 3a is a schematic diagram of the first similarity matrix and the first similarity threshold;

Figure 3b is a schematic diagram of the second similarity matrix and the second similarity threshold;

Figure 4 is a schematic flow chart of similarity threshold update for feature binding;

Figure 5 is another flow diagram of a method for improving scene adaptability of target detection;;

Figure 6 is another schematic flow chart of similarity threshold update for feature binding;

Figure 7 is a schematic diagram of a device for improving scene adaptability of target detection;

Figure 8 is another schematic diagram of a device for improving scene adaptability of target detection.

Detailed ways

In order to make the purpose, technical solution, and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and examples. Obviously, the described embodiments are only some of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art fall within the scope of protection of the present invention.

In order to make the purpose, technical means and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings.

The applicant's research found that false positives and false negatives in target detection results are particularly related to detection thresholds. The existing detection thresholds are generated by setting false positive rates, for example, based on false positive rates of 1/100000, 1/1000000, etc. Generate detection threshold. The resulting false positive rate is often based on the vendor's own test set, which has test target data in many different scenarios, but does not produce a separate false positive rate value for each usage scenario. In actual use, there is often a large gap between a fixed single usage scenario and the manufacturer's self-test environment. If the detection threshold is generated based only on the preset false alarm rate, there may be a large deviation between the detection threshold and the actual detection result. The problem.

In view of this, this application provides a method for improving scene adaptability of target detection, screening out targets with high frequency of false positives and/or false negatives, and adaptively generating and matching images for each feature in the false positive feature library and false negative feature library. The target data matches the detection threshold, thereby utilizing the feature library at a smaller granularity and improving the scene adaptability of target detection.

Refer to Figure 1, which is a schematic flow chart of a method for improving scene adaptability of target detection. The method includes: on the target detection device side,

Step S101: Obtain the detection result of the current target data.

Step S102: Extract target features from the current target data.

Step S103: Determine the matching degree between each target feature and each feature in the feature database.

Among them, the matching degree can be determined by similarity or Euclidean distance.

Step S104: Compare each matching degree with the detection threshold bound by the matching feature in the feature library used to determine the matching degree.

Step S105: Determine the processing of the detection result according to the comparison result.

in,

The feature library is a feature set of target data calibrated according to whether the detection result is a false positive or a false negative.

Specifically, the matching degree can be measured by similarity, or by parameters such as Euclidean distance. And since the matching degree is compared with the detection threshold bound to the matched feature that determines the matching degree, the form of the detection threshold should be the same as the matching degree. Therefore, the detection threshold in this application can also be regarded as a match. degree threshold. For example, when the matching degree is measured by similarity, the detection threshold may refer to the similarity threshold; when the matching degree is measured by Euclidean distance, the detection threshold may refer to the distance threshold.

The embodiments of the present application use false positives and/or missed target data to establish a feature database, and set a detection threshold for each feature in the feature database. In this way, each feature is bound to a different detection threshold, thereby improving detection efficiency. The adaptability of the threshold to the scene thereby improves the scene adaptability of the detection results obtained based on the detection threshold.

In order to facilitate understanding of this application, the following description takes a security system as an example. It should be understood that this application is not only applicable to security systems, but is also applicable to any monitoring system that utilizes target detection in industrial applications.

In a possible embodiment, the matching degree of each target feature and each feature in the feature database can be determined by calculating the similarity between each target feature and each feature in the feature database. In this embodiment, comparing each matching degree with the detection threshold bound by the matching feature in the feature library used to determine the matching degree may include: for each calculated similarity, Compare with the similarity threshold; where the similarity threshold is the similarity threshold bound to the feature used for the similarity calculation in the feature library. Determining the processing of the detection result according to the comparison result may include: correcting the detection result according to the comparison result.

Specifically, the feature library may include features used to characterize the first target data whose detection result is a false positive. The first feature library of the collection set may also include: a second feature library used to characterize the feature set of the second target data whose detection result is false negative. The correction of detection results under different circumstances will be described in detail below.

Refer to Figure 2, which is a schematic flow chart of a method for improving scene adaptability of target detection according to an embodiment of the present application. The method includes:

Step S201: Obtain the first feature database and the second feature database.

The first feature database includes a false positive feature set, which is used to characterize the first target data whose detection result is a false positive, and the second feature database includes a false negative feature set, which is used to characterize the second target data whose detection result is a false negative. Feature collection of target data.

As an example, the first feature database and the second feature database can be established as follows:

Collect target data within a set first time period and/or set a first amount, and calibrate the collected target data. For example, traverse the collected target data to determine whether the detection result of the target data contains false positives, or False negatives: identify the targets with false positives and false negatives in the target data, and record the target data with false positives and false negatives.

Based on the recorded target data of false positives and false negatives, feature extraction is performed respectively. For example, a deep learning model is used for feature extraction, and the extracted false positive features are used as features in the first feature library, and the extracted false negatives are used as features in the first feature library. Features as features in the second feature library.

For example, if a human body is missed multiple times in a scene, and the missed human body has obvious and similar characteristics, such as a human body wearing the same uniform, such as an oil worker's uniform, and always in the same special posture, such as walking half-bent, then Features are extracted from the missed human body data. The extracted missed features include clothing features, posture features and other features. These features can be stored in the second feature library.

In this embodiment, the security system can use the perimeter algorithm to filter out high-frequency false positives and false negatives detection results to obtain target data for establishing a feature database.

The above-mentioned first feature database and second feature database can be updated according to at least one of time and quantity as an update trigger condition. For example, the first feature database and second feature database are updated regularly.

Since the feature database is the result of calibration, the cleanliness of the feature data in the feature database is improved.

Step S202, obtain the current detection result of the current target data, and determine whether the detection result of the target data that triggers the alarm rule in the security system is determined to be a positive alarm or a false alarm;

If it is a positive report, execute steps S203 to S204;

If it is a false alarm, steps S205 to S206 are executed.

Step S203: Calculate the first similarity between each target feature of the current target data and each first feature in the first feature database to obtain each first similarity.

As an example, when the detection result of the target data that triggers the alarm rule in the security system is determined to be a positive alarm, feature extraction is performed on the target data corresponding to the positive alarm. For example, a deep learning algorithm is used for feature extraction, so as to obtain at least For more than one target feature, calculate the similarity between each target feature and each first feature in the first feature library to obtain each first similarity, so as to compare each target feature with each first feature in the first feature library. The first feature is matched.

Step S204: For each first similarity, determine whether the first similarity is greater than the first similarity threshold bound to the first feature used for the first similarity calculation.

If any first similarity is greater than its first similarity threshold, the detection result is determined to be a false alarm, the detection result is corrected to a false alarm, and the alarm is no longer triggered to reduce false alarms.

Otherwise, the detection result is determined to be positive and an alarm prompt is triggered.

Step S205: Calculate the second similarity between each target feature of the current target data and each second feature in the second feature database to obtain each second similarity. In order to match each target feature with each second feature in the second feature library.

Step S206: For each second similarity, determine whether the second similarity is greater than the second similarity threshold bound to the second feature used for the second similarity calculation.

If any second similarity is greater than its second similarity threshold, the detection result is corrected to a positive report and an alarm prompt is triggered. Otherwise, the detection result is determined to be a false alarm, and the alarm will no longer be triggered to reduce false alarms.

Through the above steps S204 and S206, the matching degree can be compared with the detection threshold bound to the matching feature in the feature library used to determine the matching degree, and the processing logic of the detection result can be determined based on the comparison result. In this way , when the extracted target features are highly similar to the false positive database, the alarm process of these target data can be terminated to reduce false positives; when the extracted target features are highly similar to the false negative database, the alarm process can be These target data are re-alarmed to increase the detection rate, which is beneficial to improving the scene adaptability of target detection.

Exemplarily, the target features extracted from the target feature data 1 include: target feature 1, target feature 2,..., target feature n; the first feature library includes: first feature 1, first feature 2,..., first feature i; the second feature library includes: second feature 1, second feature 2,..., second feature j; the first feature Feature 1 is bound to the similarity threshold 1, the first feature 2 is bound to the similarity threshold 2,..., the first feature i is bound to the similarity threshold i; the second feature 2 is bound to the similarity threshold 1, and the second feature 2 is bound Similarity threshold 2,..., the second feature j is bound to similarity threshold j. For example, a human body image is used as target data 1. The target features extracted from the target data 1 include eye features, facial features, and body features. The first feature in the first feature library includes eye features, facial features, and body features. etc. Among them, the eye features are bound with an eye similarity threshold of 1, the facial features are bound with a facial similarity threshold of 2, and the body features are bound with a body similarity threshold of 3. Similarly, the second feature in the second feature library Features include eye features, facial features, and body features. Among them, eye features are bound with an eye similarity threshold of 1', facial features are bound with a facial similarity threshold of 2', and body features are bound with a body similarity threshold of 3. '.

If the detection result of target feature data 1 is positive, then calculate the similarity between target feature 1 and each first feature, and obtain each first similarity. As the first similarity result, then target feature 1 has the i-th A similarity result, then there are i×n first similarity results for n target features, which can be represented by a matrix Cin with i rows and n columns, where the elements in the i-th row and n-th column of the matrix are the first features i and The first similarity of target feature n; each row in the matrix is bound to the first similarity threshold. As shown in Figure 3a, Figure 3a is a schematic diagram of the first similarity matrix and the first similarity threshold. The matrix Cin in row i and column n is the first similarity matrix, and the matrix Cin in row i and column n is the first similarity matrix. The element is the first similarity between the first feature i and the target feature n. Each row in the matrix Cin is bound to a first similarity threshold, indicating the first similarity threshold bound to each first feature. For example, row 1 is bound to the first similarity threshold 1, that is, the first feature 1 is bound to the first similarity threshold 1; row 2 is bound to the first similarity threshold 2, that is, the first feature 2 is bound to the first Similarity threshold 2; the i-th row is bound to the first similarity threshold i, that is, the first feature i is bound to the first similarity threshold i.

For each row of elements in the matrix Cin, once there is any first similarity result in the row element that is greater than the first similarity threshold bound to the row, it means that the target feature corresponding to the element is the same as the first similarity result corresponding to the element. The feature similarity is high and the matching degree is large, so the detection result is a false positive; if there is no first similarity result in the row element that is greater than the first similarity threshold bound to the row, that is, the row If the elements are all smaller than the first similarity threshold bound to the row, it means that the target feature corresponding to the element has a low similarity with the first feature, so the detection result is not a false positive.

Similarly, if the detection result of target feature data 1 is a false alarm, calculate the similarity between target feature 1 and each second feature to obtain each second similarity. As the second similarity result, then target feature 1 has j second similarity results, then there are j×n second similarity results for the n target features, It can be represented by a matrix Cjn with j rows and n columns, where the elements in the j-th row and n-th column in the matrix are the second similarity between the second feature j and the target feature n; each row in the matrix is bound to the second similarity threshold. As shown in Figure 3b, Figure 3b is a schematic diagram of the second similarity matrix and the second similarity threshold. The matrix Cjn in the jth row and nth column is the second similarity matrix. The jth row and nth column in the matrix Cjn The element is the second similarity between the second feature j and the target feature n. Each row in the matrix Cjn is bound to a second similarity threshold, indicating the second similarity threshold bound to each second feature. For example, row 1 is bound to the second similarity threshold 1, that is, the second feature 1 is bound to the second similarity threshold 1; row 2 is bound to the second similarity threshold 2, that is, the second feature 2 is bound to the second The similarity threshold is 2; the j-th row is bound to the second similarity threshold j, that is, the second feature j is bound to the second similarity threshold j.

For each row element in the matrix Cjn, once there is any second similarity result in the row element that is greater than the second similarity threshold bound to the row, it means that the target feature corresponding to the element is the same as the second similarity result corresponding to the element. The feature similarity is high and the matching degree is large, so the detection result is a false negative; if there is no second similarity result in the row element that is greater than the second similarity threshold bound to the row, that is, the row If the elements are all smaller than the second similarity threshold bound to the row, it means that the target feature corresponding to the element has a low similarity with the second feature, so the detection result is not a false negative.

The similarity thresholds bound to the features in the above feature database are generated through an adaptive method. The similarity thresholds bound to each feature in these feature databases are different, and each similarity threshold changes with the matched feature. It is adjusted based on the degree of matching with the target characteristics of the target data within the preset range, wherein the target data within the preset range includes target data within a set second time period and/or a set second quantity. Considering that the target data in the scenario changes with time, business characteristics and other factors, the similarity threshold bound to the feature in the feature library will be updated when the update triggering conditions are met.

The generation of the similarity threshold is explained below.

Refer to Figure 4, which is a schematic diagram of updating the similarity threshold of feature binding. When the update conditions are met, the update of the similarity threshold of feature binding in the feature database is triggered, where the update conditions include: meeting at least one of the set update time, the set update frequency, and the feature database being updated. The update process includes: for any feature in the feature database,

Step S401: Calculate the similarity between the feature and each target feature of all target data within the preset range. Among them, all target data are target data within a set time threshold and/or within a set quantity threshold. The target features of the target data can also be used to calibrate the detection results, so as to obtain a feature library.

Step S402: Select the maximum similarity from all similarity results of the feature.

Step S403: Adjust the similarity threshold of the feature binding based on the maximum similarity. For example, update the similarity threshold of the feature binding to be greater than or equal to the sum of the maximum similarity and the amount of redundancy.

As an example, the first similarity threshold bound to the first feature is obtained by calculating the similarity between the first feature and the target feature of the third target data, where the third target data is within the first preset range For data within which the detection result is positive, the first preset range is the set third time period and/or the set third quantity. Specifically, the first similarity threshold floats around the maximum similarity sim1_max between the first feature and the target feature, and the floating value depends on the specific correction strategy adopted. As an example, the first similarity threshold needs to satisfy thr1≥sim1_max+gap1, where thr1 is the first similarity threshold and gap1 is the first redundancy amount, which is used to represent the floating maximum similarity based on the first feature. value.

The second similarity threshold bound to the second feature is obtained by calculating the similarity between the second feature and the target feature of the fourth target data, where the fourth target data is within the second preset range and the detection result is For false alarm data, the second preset range is the set fourth time period and/or the set fourth quantity. Specifically, the second similarity threshold floats around the maximum similarity sim2_max between the target feature and the second feature, and the floating value depends on the specific correction strategy adopted. As an example, the second similarity threshold needs to satisfy thr2≥sim2_max+gap2, where thr2 is the second similarity threshold and gap2 is the second redundancy amount, used to characterize the maximum similarity based on the second feature correspondence. Floating value.

This embodiment helps improve the scene adaptability of the detection results by adaptively adjusting the detection threshold bound to each feature in the feature library, and can solve the problem of missed negatives and false positives in the existing security system, especially For repetitive omissions and false positives, the detection results can be corrected by using the matching degree of the target features of the target data and the feature library and the detection threshold comparison results, which will help improve the accuracy and reliability of the detection results and reduce the risk of False negatives and false positives.

In specific applications, the method for improving scene adaptability of target detection provided by the embodiment of the present application can be shown in Figure 5: judging whether the current target data needs to alarm based on the inherent method, where the inherent method refers to: based on the detection result of the target data It reflects the monitored status. When the detection results meet the preset conditions, for example, when the category and confidence level of the detected target reach the preset conditions, an alarm or other prompts will be triggered.

If it is, an alarm is issued for the current target data, indicating that the detection result of the current target data is a positive report, then the target features of the current target data are extracted, and the target features are compared with the typical false alarm feature library (i.e. Perform feature similarity calculation on each feature in the aforementioned first feature database) to obtain each first similarity, that is, execute the aforementioned step S203.

And for each first similarity, compare the first similarity with the first similarity threshold bound to the first feature used for the first similarity calculation, and determine whether the first similarity is greater than the first similarity used for the first similarity calculation. The first similarity threshold bound to the first feature calculated by the first similarity. If yes, then no alarm will be issued; if not, then alarm will be issued. That is, the aforementioned step S204 is executed.

If yes, that is, an alarm is issued for the current target data, indicating that the detection result of the current target data is a positive report, then the target features of the current target data are extracted, and the target features are compared with those in the typical false alarm feature library (i.e., the aforementioned first feature library) Feature similarity calculation is performed on each first feature to obtain each first similarity. And for each first similarity, compare the first similarity with the first similarity threshold bound to the first feature used for the first similarity calculation, and determine whether the first similarity is greater than the first similarity used for the first similarity calculation. The first similarity threshold bound to the first feature calculated by the first similarity; if yes, no alarm will be issued; if no, then an alarm will be issued. That is, the aforementioned steps S203 to S204 are executed.

If not, that is, no alarm is issued for the current target data, indicating that the detection result of the current target data is a false alarm, then the target features of the current target data are extracted, and the target features are compared with those in the typical false negative feature library (i.e., the aforementioned second feature library) Feature similarity calculation is performed on each second feature to obtain each second similarity. And for each second similarity, compare the second similarity with the second similarity threshold bound to the second feature used for the second similarity calculation, and determine whether the second similarity is greater than the second similarity used for the second similarity calculation. The second similarity threshold bound to the second feature calculated by the second similarity; if yes, an alarm will be issued; if not, no alarm will be issued. That is, the aforementioned steps S205 to S206 are executed.

The update method of the similarity threshold of the feature binding in the feature library can be shown in Figure 6: when the update condition is met, the update of the similarity threshold of the feature binding in the feature library is triggered, where the update condition includes: satisfying the setting At least one of the update time, the set update frequency, and the feature database being updated.

The update process includes: for any feature in the typical feature library T, calculate the similarity between the feature and each target feature of all typical positive or false positive target data in the time period P (i.e., P time length), and calculate from Select the maximum similarity among all similarity results to implement the maximum similarity (MaxSim) calculation. When the typical feature library is a typical false positive feature library (i.e., the aforementioned first feature library), calculate the similarity between the feature and each target feature of all typical positive target data within the time period P (i.e., P time length); When the typical feature library is a typical false negative feature library (i.e., the aforementioned second feature library), calculate this feature and each target feature of all typical false positive target data within the time period P (i.e., P time length) similarity.

Through the calculated maximum similarity, the similarity threshold THRt1 of the feature binding in the typical feature library T is adjusted to THRt2. At this time, THRt2 should satisfy THRt2≥MaxSim+Gap to realize the adjustment of the similarity threshold. Among them, THRt2 is the similarity threshold of feature binding in the adjusted typical feature library T, MaxSim is the maximum similarity corresponding to the features in the typical feature library, and Gap is used to represent the floating value based on the maximum similarity corresponding to the features in the typical feature library. value.

The above update process is the aforementioned steps S401 to S403.

Refer to FIG. 7 , which is a schematic diagram of a device for improving scene adaptability of target detection. The device includes:

The target detection module is used to extract target features from the current target data and obtain the detection results of the current target data;

Feature matching module, used to match each target feature with each feature in the feature library;

A threshold adjustment module, configured to adjust the detection threshold bound to each matched feature according to the matching degree between the matched feature and the target feature of the target data within the preset range;

in,

The threshold adjustment module is also configured to adjust the similarity threshold bound to each feature according to the maximum similarity corresponding to each feature, and the maximum similarity corresponding to each feature is: The maximum value of similarity between the feature and the target feature of the target data within the preset range.

The feature matching module includes:

The first feature matching submodule is used to calculate the similarity between each target feature and each first feature in the first feature library to obtain each first similarity when the detection result is positive. ;

The second feature matching submodule calculates the similarity between each target feature and each second feature in the second feature library to obtain each second similarity when the detection result is a false positive.

The comparison module includes:

The first comparison sub-module is used to compare each calculated first similarity with the first similarity Compare the first similarity threshold bound to the calculated first feature;

The second comparison sub-module is configured to compare each calculated second similarity with a second similarity threshold bound to the second feature used for the second similarity calculation.

The determination module includes:

A first determination sub-module, configured to correct the detection result as an error when any calculated first similarity is greater than the first similarity threshold bound to the first feature used for the first similarity calculation. reported, otherwise, the detection result remains positive;

The second determination sub-module is used to correct the detection result to positive when any calculated second similarity is greater than the second similarity threshold bound to the second feature used for the second similarity calculation. Otherwise, the detection result remains a false positive.

The threshold adjustment module includes:

The first similarity threshold adjustment sub-module is used to adjust the first similarity threshold bound to each first feature according to the maximum similarity corresponding to the first feature, and the first similarity threshold corresponding to the first feature is The maximum similarity is: the maximum value of the similarity between the first feature and the target feature of the third target data whose detection result is positive within the first preset range;

The second similarity threshold adjustment sub-module is used to adjust the second similarity threshold bound to each second feature according to the maximum second similarity corresponding to the second feature. The corresponding maximum similarity is: the maximum value of similarities between the second feature and the target feature of the fourth target data whose detection result is a false alarm within the second preset range.

The device also includes:

The feature database management module is used to collect target data within a set first time period and/or set a first quantity, and calibrate the collected target data;

Based on the calibrated target data, determine whether the detection results of the collected target data contain false positives or false negatives, and record the target data with false positives or false negatives;

Based on the recorded target data, feature extraction is performed, and the extracted false positive features are used as features in the first feature library, and the extracted false negative features are used as features in the second feature library.

Refer to FIG. 8 , which is another schematic diagram of a device for improving scene adaptability of target detection. The device includes a memory and a processor, the memory stores a computer program, and the processor is configured to execute the steps of the computer program to implement the method of improving scene adaptability of target detection in the present application.

The memory may include random access memory (Random Access Memory, RAM) or non-volatile memory (Non-Volatile Memory, NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located far away from the aforementioned processor.

The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (Network Processor, NP), etc.; it can also be a digital signal processor (Digital Signal Processing, DSP), special integrated Circuit (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

Embodiments of the present invention also provide a computer-readable storage medium. A computer program is stored in the storage medium. When the computer program is executed by a processor, the steps of the method for improving scene adaptability of target detection are implemented.

As for the device/network-side device/storage medium embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For relevant details, please refer to the partial description of the method embodiment.

In this document, relational terms such as first, second, etc. are used only to distinguish one entity or operation from another entity or operation and do not necessarily require or imply the existence of any such entity or operation between these entities or operations. Actual relationship or sequence. Furthermore, the terms "comprises," "comprises," or any other variation thereof are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that includes a list of elements includes not only those elements, but also those not expressly listed other elements, or elements inherent to the process, method, article or equipment. Without further limitation, an element defined by the statement "comprises a..." does not exclude the presence of additional identical elements in a process, method, article, or apparatus that includes the stated element.

The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims

A method for improving scene adaptability of target detection, which is characterized in that the method includes:

Get the detection results of the current target data;

Extract target features from the current target data;

Determine the matching degree of each target feature with each feature in the feature library;

Compare each matching degree with the detection threshold bound to the matched feature in the feature library used to determine the matching degree;

Determine the processing of the detection results based on the comparison results;

in,

The feature library is a feature set of target data calibrated according to whether the detection result is a false positive or a false negative;

The detection threshold bound to each matched feature is adjusted according to the matching degree between the matched feature and the target feature of the target data within the preset range.
The method for improving scene adaptability according to claim 1, wherein determining the matching degree between each target feature and each feature in the feature library includes:

Calculate the similarity between each target feature and each feature in the feature library;

Comparing each matching degree with the detection threshold bound to the matching feature in the feature library used to determine the matching degree includes:

For each similarity calculated;

Compare this similarity to a similarity threshold;

in,

The similarity threshold is the similarity threshold bound to the features used for the similarity calculation in the feature library.
The method for improving scene adaptability according to claim 2, characterized in that the similarity threshold bound to each feature is adjusted according to the maximum similarity corresponding to each feature, wherein each feature corresponds to The maximum similarity is: the maximum value of the similarity between each feature and the target feature of the target data within the preset range;

Determining the processing of the detection result based on the comparison result includes:

The detection results are corrected based on the comparison results.
The method for improving scene adaptability according to claim 3, characterized in that the detection results include positive reports, and the feature library includes: a first feature library used to characterize the first target data whose detection results are false positives. feature set;

The calculation of the similarity between each target feature and each feature in the feature library includes:

In the case where the detection result is a positive report, calculate the similarity between each target feature and each first feature in the first feature library to obtain each first similarity;

The correction of the detection results according to the comparison results includes:

If any calculated first similarity is greater than the first similarity threshold bound to the first feature used for the first similarity calculation, the detection result is corrected as a false positive;

in,

The first similarity threshold bound to each first feature is adjusted according to the maximum similarity corresponding to the first feature. The maximum similarity corresponding to the first feature is: the first feature and the first The maximum value of the similarity between the target features of the third target data whose detection result is a positive report within the preset range.
The method for improving scene adaptability according to claim 4, wherein the first similarity threshold bound to each first feature is adjusted according to the maximum similarity corresponding to the first feature, including :

It is determined that the first similarity threshold is greater than the sum of the maximum similarity corresponding to the first feature and a first redundancy amount, where the first redundancy amount is used to characterize the maximum similarity based on the first feature correspondence. Floating value.
The method for improving scene adaptability according to claim 3, characterized in that the detection results include false alarms,

The feature library includes: a second feature library, a feature set used to characterize the second target data whose detection results are missed;

The calculation of the similarity between each target feature and each feature in the feature library includes:

In the case where the detection result is a false positive, calculate the similarity between each target feature and each second feature in the second feature library to obtain each second similarity;

The correction of the detection results according to the comparison results includes:

If any calculated second similarity is greater than the second similarity threshold bound to the second feature used for the second similarity calculation, then the detection result is corrected to a positive report;

in,

The second similarity threshold bound to each second feature is adjusted according to the maximum second similarity corresponding to the second feature. The maximum similarity corresponding to the second feature is: the second feature and No. The maximum value among the similarities between the target features of the fourth target data whose detection result is a false alarm within the two preset ranges.
The method for improving scene adaptability according to claim 6, wherein the second similarity threshold bound to each second feature is adjusted according to the maximum second similarity corresponding to the second feature. ,include:

It is determined that the second similarity threshold is greater than the sum of the maximum similarity corresponding to the second feature and a second redundancy amount, wherein the second redundancy amount is used to characterize the maximum similarity based on the second feature correspondence. Floating value.
The method for improving scene adaptability according to claim 1, characterized in that the feature library is established as follows:

Collect target data within a set first time period and/or set a first quantity, and calibrate the collected target data;

Based on the calibrated target data, determine whether the detection results of the collected target data contain false positives or false negatives, and record the target data with false positives or false negatives;

Based on the recorded target data, perform feature extraction, use the extracted false positive features as features in the first feature library, and use the extracted false negative features as features in the second feature library;

Extracting target features from the current target data includes: extracting target features from the target data whose detection results do not meet the set conditions;

The similarity threshold is triggered to be updated according to the set update conditions;

The preset range includes a set second time period and/or a set second number;

The scene described is the scene monitored by the security system.
A device for improving scene adaptability of target detection, characterized by:

The target detection module is used to obtain the detection results of the current target data and extract target features from the current target data;

A feature matching module, used to match each target feature with each feature in the feature library, where the feature library is a feature set of target data calibrated according to whether the detection result is a false positive or a false negative;

A comparison module, used to compare the detection results with the detection threshold bound to the matching feature in the feature library;

A determination module, used to determine the processing of the detection result according to the comparison result;

The threshold adjustment module is used to adjust the detection threshold bound to each of the matched features. Adjusted based on the matching degree between the matched feature and the target feature of the target data within the preset range.
A target detection system, characterized by comprising the device for improving the scene adaptability of target detection as claimed in claim 9.