False alarm reduction method and system for multi-task target detection

By adding a feature extraction branch and a false alarm library to the target detection model for potential hazards in power transmission channels, and using a joint feature similarity function to filter false alarms, the problem of high false alarm rate caused by insufficient model generalization ability is solved, and the effect of efficiently reducing the false alarm rate on terminal equipment with limited resources is achieved.

CN121033737APending Publication Date: 2025-11-28JINAN XINTONG ELECTRIC TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410625124.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-05-20
Publication Date
2025-11-28

AI Technical Summary

Technical Problem

Existing technologies suffer from high false alarm rates in detecting potential hazards in power transmission channels, especially when the model's generalization ability is limited, leading to an increased false alarm rate. Furthermore, existing solutions are costly or resource-intensive.

Method used

In the hazard target detection model, a feature extraction branch is added to construct a false alarm database. By calculating the similarity between the target feature vector and the false alarm database through a joint feature similarity function, false alarm targets are filtered out, thereby reducing the false alarm rate.

Benefits of technology

Without increasing model complexity and resource consumption, it effectively reduces the false alarm rate, improves detection accuracy, and is versatile, applicable to fields such as face recognition and vehicle recognition.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121033737A_ABST
    Figure CN121033737A_ABST
Patent Text Reader

Abstract

The invention discloses a false alarm reduction method and system for multi-task target detection, and the method comprises the steps: inputting obtained image data into a trained hidden danger target detection model, and outputting a hidden danger target detection result and a target feature vector; if a hidden danger target is detected and a camera ID corresponding to the image data is in a pre-constructed false alarm library, respectively calculating the similarity between the target feature vector and the target hash value and the feature vector and the hash value of a false alarm target under the same camera ID in the false alarm library to obtain a joint feature similarity; and based on the value of the joint feature similarity, in combination with the matching degree of the target frame IoU, judging whether the detected hidden danger target is a false alarm target, if so, filtering out an alarm for the hidden danger target, and if not, normally alarming. On the premise that the model recall rate is not reduced, it is guaranteed that the found misinformation target is a real misinformation target, and the misinformation rate is reduced; the method can be applied to various fields such as power transmission channel hidden danger recognition, face recognition or vehicle recognition.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a method and system for reducing false alarms in multi-task target detection. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In the field of machine vision, research on target detection using computer image processing technology is becoming increasingly popular. It is closely related to many real-world applications. For example, accurate identification of potential hazards along power transmission lines (such as tower cranes, smoke, cranes, and wildfires) ensures accurate and reliable power grid operation alarms (e.g., misidentifying windmill blades as cranes could trigger false alarms). In passenger flow statistics and analysis, accurate facial recognition can precisely profile users and assist businesses in providing customized services. However, when deployed, facial detection models may exhibit false positives due to limitations in generalization ability (e.g., misidentifying faces on billboards as real faces), leading to errors in passenger flow statistics and analysis. In the field of intelligent transportation, vehicle detection, license plate analysis, and vehicle attribute analysis assist traffic management departments in big data analysis. However, after deployment, the limited generalization ability of the models can also lead to false positives (e.g., misidentifying vehicle symbols on road signs as vehicles), resulting in unreliable big data analysis results.

[0004] Taking the identification of hidden dangers in power transmission channels as an example, current technologies mostly rely on video or image monitoring devices for online target monitoring and early warning. Among these, the use of deep learning technology for the detection and identification of hidden danger targets has been widely applied. Deep learning technology trains models on large amounts of data, enabling them to learn the features of target objects and thus achieve automatic detection. However, using deep learning technology for the detection of hidden danger targets in power transmission channels still faces the following technical challenges:

[0005] (1) Due to the limited image data of existing potential hazards, the environments and scenarios encountered when deploying neural network models are highly complex for various reasons. This limits the model's generalization ability, leading to false alarms in unfamiliar scenarios (e.g., mistaking windmill blades for cranes), thus reducing the model's practicality. To reduce the false alarm rate, existing technologies mainly focus on the following two aspects:

[0006] ① This can be addressed through model training. This can be done by modifying the network structure to increase the model's recognition accuracy, or by adding data samples from unseen scenes for training, thus increasing the model's generalization ability to unfamiliar scenarios. However, this approach is time-consuming and costly. Modifying the network structure requires extensive training and validation to ensure model effectiveness, and adding labeled training datasets also incurs significant manpower and time costs.

[0007] ② By introducing technologies such as multi-sensor fusion, more accurate target detection can be achieved, thereby reducing false alarms. However, this approach also increases sensor costs, and the specific effectiveness still needs to be verified through significant experimental and time-consuming research.

[0008] (2) The neural network model structure used for target detection in the prior art is relatively complex. Although it can be deployed at the front end of the model, it is limited by the computing and storage resources at the front end. The neural network model function is often relatively simple and usually cannot add other functions to the neural network structure. Summary of the Invention

[0009] To address the aforementioned issues, this invention proposes a method and system for reducing false alarms in multi-task target detection. A feature extraction branch is added to the model's output, enabling feature extraction with fewer model parameters and achieving lightweight front-end deployment. Simultaneously, a false alarm database is constructed, and the similarity between the target feature vector and the feature vectors in the database is calculated based on a joint feature similarity function, thereby filtering out false alarms and improving the accuracy of target detection.

[0010] In some implementations, the following technical solutions are adopted:

[0011] A method for reducing false alarms in multi-task target detection includes:

[0012] Acquire monitoring image data, input the acquired image data into the trained hidden danger target detection model, and output the hidden danger target detection results and target feature vectors;

[0013] If a potential hazard is detected, it is determined whether the camera ID corresponding to the image data is in a pre-built false alarm database.

[0014] If the target is in the false alarm database, calculate the similarity between the target feature vector and target Phash value and the feature vector and Phash value of the false alarm target with the same camera ID in the false alarm database to obtain the joint feature similarity; based on the value of the joint feature similarity and the matching degree of the target bounding box IoU, determine whether the detected potential target is a false alarm target. If so, filter out the alarm for the potential target; if not, issue a normal alarm.

[0015] If it is not in the false alarm database, a normal alarm will be triggered.

[0016] The detection head of the hazard target detection model includes a feature extraction branch, a category prediction branch, and a bounding box regression branch set in parallel. The feature extraction branch is used to output the feature vector of the detected target, the category prediction branch is used to predict the classification result of the detected target, and the bounding box regression branch is used to output the position of the target box.

[0017] The loss function of the hidden danger target detection model is as follows:

[0018] L det =L m +L GIoU +L cls +L reg

[0019] in,

[0020] f a f p f n These are the feature vectors output by the feature extraction branch, f and f respectively. a f is an eigenvector of any type. n with f a f is the feature vector of the negative sample with inconsistent prediction type. p with f a The feature vectors of positive samples with consistent prediction types are given, where m is the dynamically scaling threshold, 0. <m<1;L GIoU For GIoU loss, L cls For classification loss, L reg This represents the L1 regression loss.

[0021] As a further solution, the false alarm database stores the feature vector, Phash value, target detection box, and target type corresponding to each false alarm target under each camera ID.

[0022] As a further solution, the joint feature similarity is specifically as follows:

[0023]

[0024] Where β is the weighting factor in the combined score, S c S represents the similarity between the target feature vector and the feature vectors of the same camera ID in the false alarm database; p The similarity between the target Phash value and the Phash value corresponding to the feature vector under the same camera ID in the false alarm database.

[0025] In other embodiments, the following technical solutions are adopted:

[0026] A false alarm reduction system for multi-task target detection includes:

[0027] The data acquisition module is configured to acquire surveillance image data;

[0028] The hazard target detection module is configured to input the acquired image data into the trained hazard target detection model and output the hazard target detection result and the target feature vector.

[0029] The joint feature similarity calculation module is configured to calculate the similarity between the target feature vector and the target Phash value and the feature vector and Phash value of the falsely reported target under the same camera ID in the false report database when a potential target is detected and the camera ID corresponding to the image data is in the pre-built false report database, so as to obtain the joint feature similarity.

[0030] The false alarm target filtering module is configured to determine whether a detected potential hazard target is a false alarm target based on the value of joint feature similarity and the matching degree of the target bounding box IoU. If it is, the alarm for the potential hazard target is filtered out; otherwise, a normal alarm is issued.

[0031] In other embodiments, the following technical solutions are adopted:

[0032] A terminal device includes a processor and a memory, wherein the processor implements instructions; and the memory stores multiple instructions adapted to be loaded by the processor and executed by the processor to reduce false alarms in the multi-task target detection method described above.

[0033] In other embodiments, the following technical solutions are adopted:

[0034] A computer-readable storage medium storing a plurality of instructions adapted to be loaded by a processor of a terminal device and executed by the aforementioned method for reducing false alarms in multi-task target detection.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] (1) This invention calculates the joint feature similarity between the detected target hazard and the false alarm target under the same camera ID in the false alarm database. The joint feature similarity introduces the Phash value, which can enhance the accuracy of similarity comparison. At the same time, combined with the IoU target box matching strategy and the idea that the target type and camera ID are consistent, it can ensure that the false alarm target found is the real false alarm target without reducing the model recall rate, thus reducing the false detection probability of the model.

[0037] (2) The present invention adds a feature extraction branch to the detection head of the hidden danger target detection model, and constructs a feature extraction module with a very small number of parameters. Based on the idea of ​​multi-task, it can simultaneously realize tasks such as target detection, feature extraction and target box position detection; there is no need to build a separate feature extraction model, which ensures that the false alarm reduction scheme of the present invention can be effectively deployed on terminal devices with limited resources, and effectively reduces the false alarm rate without reducing the recall rate of the algorithm.

[0038] (3) The false alarm reduction method of the present invention has good versatility and can be effectively extended to other technical fields such as face recognition or vehicle recognition that need to reduce the false alarm rate of target detection, reduce the fixed false alarm targets predicted by the model, and improve the accuracy of the deployed model.

[0039] Other features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0040] Figure 1 This is a flowchart of the false alarm reduction method for multi-task target detection in an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the hidden danger target detection model structure in an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram of the identification results of the hidden danger target detection model in an embodiment of the present invention. Detailed Implementation

[0043] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0044] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0045] Terminology Explanation:

[0046] Phash: Perceptual hashing algorithm;

[0047] IoU: Intersection over Union (IoU) ratio;

[0048] GIoU: Generalized IoU, is an improved method of IoU calculation.

[0049] Example 1

[0050] In one or more embodiments, a method for reducing false alarms in multi-task target detection is disclosed, combining... Figure 1 Specifically, it includes the following processes:

[0051] S101: Acquire monitoring image data of the transmission line channel, input the acquired image data into the trained hidden danger target detection model, and output the hidden danger target detection results and target feature vectors.

[0052] In this embodiment, the structure of the hidden danger target detection model is as follows: Figure 2 As shown, the process includes: input image data sequentially passing through a backbone network and a neck structure before entering the head (detection head). The backbone uses a CSPneXt structure, with the backbone constructed from CSP-blocks, yielding C3, C4, and C5 feature maps as output. The neck uses a CSPneXtPAFPN structure, applying a 1x1 convolution to C5 with 512 channels to obtain P5. P5 is then upsampled to obtain P4. P4 and C4 are concatenated. P4 is then upsampled again to obtain P3, which is also concatenated with C3. P3 undergoes two downsampling operations to obtain P4 and P5, with P4 and P5 having the same feature dimensions as P3. P3, P4, and P5 are then concatenated. This process promotes inter-layer feature interaction, achieving the effect of extracting rich feature details across multiple levels. Finally, a 3x3 convolution kernel is used, and a convolution operation with a stride of 1 transforms the final output feature into 256.

[0053] The main improvement of the hidden danger target detection model in this embodiment is that, based on the idea of ​​multi-task, a feature extraction branch is added to the head part to build a lightweight feature extraction module with fewer model parameters, so that a separate neural network model for feature extraction is no longer needed, which enables the model to be deployed on terminal devices with limited resources.

[0054] The head section of this embodiment includes three branches: a category prediction branch, which predicts the classification result of the detected target, and reduces the number of channels from 256 to 10, with 10 corresponding to the total number of categories; a feature extraction branch, which outputs the feature vector of the detected target, and reduces the number of channels from 256 to 64, with 64 representing the output dimension of the feature vector; and a bounding box regression branch, which outputs the position of the target bounding box, and reduces the number of channels from 256 to 4, with 4 representing the coordinates of the bounding box.

[0055] Meanwhile, this embodiment reconstructs the target detection loss function L. det, so that the feature extraction module is trained while training the object detection. The loss function L det Specifically:

[0056] L det = L m + L GIoU + L cls + L reg

[0057] Among them, f a f n f p These three vectors are 64-dimensional feature vectors of any object output by the feature extraction module. f a is a feature vector of any type, and f n is the feature vector of any object whose prediction type is inconsistent with f a . f p is the feature vector of any object whose prediction type is consistent with f a . k represents the k-th sample pair in the sample pairs with inconsistent types in batch training, K is the total number of sample pairs with inconsistent prediction types in batch training, r represents the r-th sample pair in the sample pairs with consistent batch types, and R is the total number of sample pairs with consistent prediction types in batch training.

[0058] As a specific example, a sample pair with consistent types refers to a sample pair of the same smoke target, and a sample pair with inconsistent types refers to a sample pair of a smoke target and a wildfire target; in this example, f a can be the feature vector expression of the smoke target, f p can be the feature vector expression of another smoke, and f n can be the feature vector expression of the wildfire.

[0059] m is the dynamic scaling threshold, 0 < m < 1. Its value represents the size of the discrimination boundary between different categories. The larger the m value, the larger the discrimination boundary of different types of features; the smaller the m value, the smaller the discrimination boundary of different types of features. In this patent, it is set to 0.5 to retain a certain margin and ensure the normal convergence of training.

[0060] The classification loss L cls adopts qualityfocalloss, L<​​​​​A training sample set was constructed to train the hazard target detection model. Historical monitoring data of each camera in the power transmission line channel were obtained. Each data point was manually labeled with the hazard type, which included 10 types of targets: tower crane, smoke, crane, wildfire, cement pump truck, excavator, pile driver, bulldozer, foreign object in the conductor, and bird damage. The labeling content included the type of hazard target and the target bounding box.

[0062] Use these labeled data to construct a training sample set and train the hazard target detection model.

[0063] S102: If a potential target is detected, determine whether the camera ID corresponding to the image data is in the pre-built false alarm database. If it is, calculate the similarity between the target feature vector and target Phash value and the feature vector and Phash value of the false alarm target with the same camera ID in the false alarm database to obtain the joint feature similarity. If it is not, proceed with the normal alarm process.

[0064] Specifically, in combination Figure 1 In this embodiment, {u1,u2,...,u} are obtained. n Multiple days (non-continuous) of real-time monitoring images were collected, with a minimum of four days (one day per quarter). The real-time image data was then used to perform forward inference using a hazard target detection model to obtain the target type, target feature vector, and target detection bounding box for each image. The results were then analyzed using a method that calculates the bounding box every two days. i ,u j Images with the same camera ID are subjected to IoU (Intersection over Union) matching of their bounding boxes. Targets with an IoU greater than 0.5 and the same target type are identified as false positives. If necessary, false positives need to be extracted from the original image for secondary manual review to ensure the accuracy of the false positive database.

[0065] A false alarm database is built based on the identified false alarm targets. The database stores the feature vector, Phash value, target detection box, and target type corresponding to each false alarm target under each camera ID. It is stored in a dictionary format: {“camera_id”:[list of feature vectors, list of Phash values, list of detection boxes, list of types]}.

[0066] If any image to be detected is found to contain a potential hazard by the hazard detection model, it is determined whether the camera image ID of the image to be detected is in the false alarm database. If not, the hazard alarm is triggered normally. If it is, the detection results of the image to be detected are retrieved, including: target feature vector, target category, target detection box, and target Phash value.

[0067] Based on the retrieved data, the similarity between the target's feature vector and the feature vectors of falsely reported targets with the same camera ID in the false alarm database is calculated:

[0068]

[0069] Where i represents any dimension of the vector. n is the total length of the vector, which is 64 dimensions in this scheme, and x i y represents the value of the i-th dimension of the target feature vector. i This represents the value of the i-th dimension of the feature vector of a false alarm target Q under the same camera ID in the false alarm database.

[0070] Calculate the similarity between the target's hash value and the hash values ​​of falsely reported targets Q under the same camera ID in the false alarm database:

[0071]

[0072] Based on the two similarities calculated above, calculate the joint feature similarity S. d :

[0073]

[0074] Where β is the weighting factor in the combination score, and the magnitude of β determines S c S p The greater the β value, the more important S is. p A small β indicates that S is valued. c This patent tends to favor S. c Therefore, β is set to 0.5.

[0075] S103: When the camera ID corresponding to the image data is in the pre-built false alarm library, based on the value of joint feature similarity and the matching degree of the target box IoU, it is determined whether the detected hidden danger target is a false alarm target. If so, the alarm for the hidden danger target is filtered out; otherwise, a normal alarm is issued.

[0076] Specifically, the joint feature similarity S between the detected target and all falsely reported targets under the same camera ID is calculated according to the process in S102. d Then, the following conditions are determined:

[0077] ① The calculated joint feature similarity S d The maximum value is greater than the set first threshold;

[0078] ② Joint feature similarity S d The maximum value corresponds to the target bounding box of the false alarm target in the false alarm library, and the IoU matching result of the target bounding box of the detected hidden danger target is greater than the set second threshold.

[0079] ③ Joint feature similarity S dThe type of the false alarm target corresponding to the maximum value in the false alarm database is consistent with the type of the detected potential hazard target;

[0080] If all three conditions ①②③ above are met simultaneously, the detected potential hazard is considered a false alarm and the alarm for that target is filtered out; if they are not met simultaneously, a normal alarm is triggered.

[0081] The first threshold and the second threshold are selected according to actual needs. They can be the same or different. For example, in this embodiment, the first threshold is set to 0.5. In addition, considering that the camera may be deflected by wind, the second threshold can also be set to 0.5.

[0082] Figure 3 A schematic diagram of the detection results obtained using the method of this embodiment is given, in which green targets are the retained detection structures, and red targets are false alarm targets that successfully match the false alarm library and need to be filtered.

[0083] This embodiment introduces a Phash value to construct joint feature similarity, which can enhance the accuracy of similarity comparison. At the same time, combined with the IoU target box matching strategy and the idea that the target type and camera ID are consistent, it can ensure that the false positive targets found are real false positive targets without reducing recall, thereby reducing the false positive rate.

[0084] The above implementation process is used as an example of identifying hidden dangers in power transmission channels to illustrate the method for reducing false alarms in multi-task target detection of the present invention. Of course, the method of the present invention can also be applied to other technical fields, such as face recognition, vehicle recognition, etc. The specific implementation method is the same as the process of identifying hidden dangers in power transmission channels, only the image data acquired and the target extracted from the acquired image data are different.

[0085] Example 2

[0086] In one or more embodiments, a false alarm reduction system for multi-task target detection is disclosed, comprising:

[0087] The data acquisition module is configured to acquire surveillance image data;

[0088] The hazard target detection module is configured to input the acquired image data into the trained hazard target detection model and output the hazard target detection result and the target feature vector.

[0089] The joint feature similarity calculation module is configured to calculate the similarity between the target feature vector and the target Phash value and the feature vector and Phash value of the falsely reported target under the same camera ID in the false report database when a potential target is detected and the camera ID corresponding to the image data is in the pre-built false report database, so as to obtain the joint feature similarity.

[0090] The false alarm target filtering module is configured to determine whether a detected potential hazard target is a false alarm target based on the value of joint feature similarity and the matching degree of the target bounding box IoU. If it is, the alarm for the potential hazard target is filtered out; otherwise, a normal alarm is issued.

[0091] It should be noted that the specific implementation methods of the above modules have been described in detail in Implementation Example 1, and will not be repeated here.

[0092] Example 3

[0093] In one or more embodiments, a terminal device is disclosed, including a server. The server includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the false alarm reduction method for multi-task target detection in Embodiment 1. For simplicity, further details are omitted here.

[0094] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0095] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0096] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0097] Example 4

[0098] In one or more embodiments, a computer-readable storage medium is disclosed, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and executed to reduce false alarms in the multi-task target detection method described in Embodiment 1.

[0099] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for reducing false alarms in multi-task target detection, characterized in that, include: Acquire monitoring image data, input the acquired image data into the trained hidden danger target detection model, and output the hidden danger target detection results and target feature vectors; If a potential hazard is detected, it is determined whether the camera ID corresponding to the image data is in a pre-built false alarm database. If the target is in the false alarm database, calculate the similarity between the target feature vector and target Phash value and the feature vector and Phash value of the false alarm target with the same camera ID in the false alarm database to obtain the joint feature similarity; based on the value of the joint feature similarity and the matching degree of the target bounding box IoU, determine whether the detected potential target is a false alarm target. If so, filter out the alarm for the potential target; if not, issue a normal alarm. If it is not in the false alarm database, a normal alarm will be triggered.

2. The method for reducing false alarms in multi-task target detection as described in claim 1, characterized in that, The detection head of the hazard target detection model includes a feature extraction branch, a category prediction branch, and a bounding box regression branch arranged in parallel. The feature extraction branch is used to output the feature vector of the detected target, the category prediction branch is used to predict the classification result of the detected target, and the bounding box regression branch is used to output the position of the target box.

3. A method for reducing false alarms in multi-task target detection as described in claim 1 or 2, characterized in that, The loss function of the hidden danger target detection model is as follows: L det L m +L GIoU +L cls +L reg in, f a f p f n These are the feature vectors output by the feature extraction branch, f and f respectively. a f is an eigenvector of any type. n with f a f is the feature vector of the negative sample with inconsistent prediction type. p with f a The feature vectors of positive samples with consistent prediction types are given, where m is the dynamically scaling threshold, 0. <m<1;L GIoU For GIoU loss, L cls For classification loss, L reg This represents the L1 regression loss.

4. The method for reducing false alarms in multi-task target detection as described in claim 1, characterized in that, The false alarm database stores the feature vector, Phash value, target detection box, and target type corresponding to each false alarm target under each camera ID.

5. The method for reducing false alarms in multi-task target detection as described in claim 1, characterized in that, The joint feature similarity is specifically as follows: Where β is the weighting factor in the combined score, S c S represents the similarity between the target feature vector and the feature vectors of the same camera ID in the false alarm database; p The similarity between the target Phash value and the Phash value corresponding to the feature vector under the same camera ID in the false alarm database.

6. The method for reducing false alarms in multi-task target detection as described in claim 5, characterized in that, Similarity S p Specifically: Where t is the Phash encoding length, p1 represents the target Phash value, and p2 represents the Phash value corresponding to the feature vector under the same camera ID in the false alarm database.

7. The method for reducing false alarms in multi-task target detection as described in claim 1, characterized in that, Based on the joint feature similarity value and the matching degree of the target bounding box IoU, it is determined whether the detected potential hazard target is a false alarm target. Specifically: ① The maximum value of the calculated joint feature similarity is greater than the set first threshold; ② The target bounding box of the false alarm target in the false alarm library corresponding to the maximum value of the joint feature similarity is greater than the IoU matching result of the target bounding box of the detected hidden danger target than the set second threshold; ③ The type of the false alarm target in the false alarm database corresponding to the maximum value of the joint feature similarity is consistent with the type of the detected hidden danger target; If all three conditions are met, the detected potential hazard is considered a false alarm; if not, a normal alarm is triggered.

8. A false alarm reduction system for multi-task target detection, characterized in that, include: The data acquisition module is configured to acquire surveillance image data; The hazard target detection module is configured to input the acquired image data into the trained hazard target detection model and output the hazard target detection result and the target feature vector. The joint feature similarity calculation module is configured to calculate the similarity between the target feature vector and the target Phash value and the feature vector and Phash value of the falsely reported target under the same camera ID in the false report database when a potential target is detected and the camera ID corresponding to the image data is in the pre-built false report database, so as to obtain the joint feature similarity. The false alarm target filtering module is configured to determine whether a detected potential hazard target is a false alarm target based on the value of joint feature similarity and the matching degree of the target bounding box IoU. If it is, the alarm for the potential hazard target is filtered out; otherwise, a normal alarm is issued.

9. A terminal device comprising a processor and a memory, the processor for implementing instructions; the memory for storing multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed as the multi-task target detection false alarm reduction method according to any one of claims 1-7.

10. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded by the processor of the terminal device and executed by the method for reducing false alarms in multi-task target detection according to any one of claims 1-7.