A regional security system and method integrating target detection

Abstract

The present invention discloses a regional security system and method integrating target detection, belonging to the field of regional security technology. The system includes a control unit and several monitoring mechanisms arranged on one side of the security area; the first monitoring mechanism pre-detects the target through a sensor to obtain first target reconnaissance information; the second monitoring mechanism includes an infrared and visible light camera and an image processing unit, which performs secondary detection on the target by responding to an activation instruction to obtain second target reconnaissance information; the control unit determines whether to activate the infrared and visible light cameras for secondary detection according to the first target reconnaissance information and preset rules, and if so, controls the execution of the security strategy according to the target reconnaissance information; the image processing unit aligns, extracts and fuses the features of the collected infrared and visible light images through an improved YOLOv5s target detection network. The present invention significantly improves the accuracy and reliability of target detection by integrating multiple sensor data, and the regional security effect is effectively guaranteed.

Description

Technical Field

[0001] The invention relates to a regional security system and method integrating target detection, belonging to the technical field of regional security. Background Art

[0002] Regional security technology is a comprehensive security system designed to protect the outer boundaries of buildings or venues and define the scope of the area. It integrates a variety of sensors, monitoring tools and protective facilities to monitor and warn in real time and repel potential intrusions.

[0003] Traditional security systems can be divided into three categories: the first category is reactive regional security systems, which trigger alarms when intrusion occurs and intercept with the help of physical fences. However, in the process of physical prevention, not only will there be fire risks due to arc discharge, but the specific positioning accuracy of personnel crossing the boundary is also unsatisfactory; the second category is active regional security systems, which focus on quickly issuing sound and light alarms and voice expulsion when intrusion occurs, and do not have physical barrier functions; the third category mainly relies on security personnel to maintain security through regular and fixed-point patrols. This approach has problems such as high pressure on personnel scheduling, incomplete monitoring coverage, and delayed detection of safety hazards. The existing security measures mentioned above often operate in isolation, which is not only difficult to make comprehensive judgments based on complex intrusion situations in reality, but also faces the following challenges:

[0004] The long response time of the security early warning system increases the risk of data loss, and the impact of security incidents is easily expanded. It is difficult to ensure the security of security data during real-time transmission, and direct local storage will result in a large amount of redundant data. The sensor nodes run continuously for a long time, and the energy consumption is too high, which may cause the node to fail and increase maintenance costs.

[0005] In addition, as a model of security technology, video surveillance systems are widely used in many industries. However, once deployed, traditional video surveillance systems are limited by the fixed position of the camera and can only provide monitoring for limited spaces, which inevitably leads to blind spots. Secondly, when facing an incident that has already occurred, users need to repeatedly search through the massive video materials captured by many cameras and manually select the key clips of related events to reproduce the full picture of the incident; and for possible future police incidents, users have to frequently patrol the live images provided by each camera in real time to obtain first-hand intelligence in a timely manner. Obviously, this requires complicated manual operations, resulting in a relatively inefficient and laborious user experience.

[0006] In addition, the most critical thing is that in bad weather or when the target is far away, there will be defects such as blurred images, errors in naked eye observation, and inaccurate monitoring information, making it difficult to effectively identify and accurately respond to intrusion incidents. Therefore, the various defects of current regional security technology are prominent, and it is not enough to fully meet the public's ever-increasing safety protection needs. Summary of the invention

[0007] The purpose of the present invention is to provide a regional security system and method integrating target detection, aiming to achieve all-weather and all-round accurate monitoring of the periphery of the security area by utilizing multimodal sensor technology, so as to improve the regional security effect of buildings or places.

[0008] In order to solve the above technical problems, the present invention is implemented by adopting the following technical solutions:

[0009] On the one hand, the present invention provides a regional security system integrating target detection, which includes: a control unit, and a plurality of monitoring mechanisms arranged on one side of a security area; the monitoring mechanisms include: a fixed rod, and a first monitoring mechanism and a second monitoring mechanism arranged at the middle and top of the fixed rod;

[0010] The first monitoring mechanism includes a plurality of sensors for pre-detecting moving targets and transmitting first target reconnaissance information corresponding to the plurality of sensors detected to the control unit;

[0011] The second monitoring mechanism includes: infrared and visible light cameras and an image processing unit, which are used to respond to the activation instruction output by the control unit, perform secondary detection on the moving target, and transmit the corresponding second target reconnaissance information to the control unit;

[0012] The control unit is respectively connected to the first monitoring mechanism and the second monitoring mechanism for determining whether it is necessary to activate the infrared and visible light cameras for secondary detection according to the first target reconnaissance information and corresponding preset rules, and if so, controlling the execution of the preset security strategy according to the first target reconnaissance information and the second target reconnaissance information obtained in sequence;

[0013] The image processing unit performs feature alignment, extraction and fusion processing on the infrared images and visible light images collected by the infrared and visible light cameras through an improved YOLOv5s target detection network.

[0014] The present invention realizes rapid and accurate monitoring of security area targets through preliminary screening by the first monitoring mechanism, in-depth detection by the second monitoring mechanism, and integration of target detection technology. The system not only improves the efficiency and accuracy of security monitoring, but also reduces system resource consumption and false alarm rate, providing strong technical support for regional security.

[0015] Optionally, the first monitoring mechanism includes: an infrared sensor, a long-range millimeter-wave radar, a short-range millimeter-wave radar and an ultrasonic sensor, and the signal output ends of the infrared sensor, the long-range millimeter-wave radar, the short-range millimeter-wave radar and the ultrasonic sensor are all connected to the control unit.

[0016] Optionally, if at least two of the infrared sensor, the long-range millimeter-wave radar, the short-range millimeter-wave radar and the ultrasonic sensor simultaneously meet one of the preset rules, the infrared and visible light cameras are activated for secondary detection.

[0017] Optionally, the preset rules include: the moving speed of the active target exceeds 3 meters per second; the shape and size of the active target match the preset "human" or "vehicle" model by more than 80%, and the size of the active target is between 0.5 square meters and 50 square meters; the active target stays in the preset sensitive area for more than 60 seconds; and the moving trajectory of the active target deviates from the preset safe path by more than 10 meters or crosses the preset virtual boundary twice in a row.

[0018] Optionally, at night or under low light conditions, the threshold of the moving speed is set to 1 meter per second, and the matching degree is set to 90%.

[0019] Optionally, the monitoring mechanism further includes: a single-spherical lightning rod antenna, an audible and visual alarm, and a power supply unit for providing power for monitoring operation;

[0020] The single-spherical lightning rod antenna is installed on the top of the fixing rod and is used to transmit the lightning current after the lightning is connected to the ground through the internal metal inner conductor and the down conductor.

[0021] Optionally, the control end of the sound and light alarm is connected to the control unit, and the control unit controls the operation of the sound and light alarm according to the first target reconnaissance information and the second target reconnaissance information.

[0022] Optionally, the power supply unit includes: a power generation module and a battery pack; the power generation module is installed at the top end of the fixing rod, and its power output end is connected to the battery pack.

[0023] Optionally, the power generation module includes: one or more of a wind power generation module and a solar power generation module.

[0024] Optionally, the regional security system based on fusion target detection also includes a remote hot backup unit, which is communicatively connected to the control unit and is used to synchronize relevant data stored in the control unit.

[0025] In a second aspect, the present invention further provides a regional security method for fusion target detection applicable to the regional security system for fusion target detection according to the first aspect, comprising:

[0026] Acquire first target reconnaissance information obtained by the first monitoring agency through pre-detection;

[0027] According to the first target reconnaissance information and the corresponding preset rules, determine whether it is necessary to activate the second monitoring mechanism for secondary detection; if so, obtain the second target reconnaissance information obtained by the second monitoring mechanism after the secondary detection;

[0028] Based on the first target reconnaissance information and the second target reconnaissance information, completing the execution of the preset security strategy;

[0029] The processing of the second target reconnaissance information includes:

[0030] The infrared image and visible light image collected by the second monitoring mechanism are input into the improved YOLOv5s target detection network for feature alignment, extraction and fusion processing to obtain the second target reconnaissance information.

[0031] The present invention utilizes a first monitoring mechanism (usually including a series of non-visual sensors) deployed in a security area to perform preliminary target detection, preliminarily screen out possible security threats, and generate first target reconnaissance information. The control unit then quickly analyzes the received information to determine whether the target indicated by the information meets the conditions for further detection. If so, it indicates that the target has a high potential risk and needs to activate a higher-level monitoring device for secondary confirmation. After receiving the activation instruction, the second monitoring mechanism uses an improved YOLOv5s target detection network to accurately identify the target to generate second target reconnaissance information. Finally, the control unit comprehensively evaluates the threat level of the target and executes the preset security strategy accordingly.

[0032] Optionally, the improved YOLOv5s target detection network includes: a feature alignment unit, a multi-level feature enhancement focusing unit, and an infrared and visible light feature fusion unit, which are used to align, extract and fuse features of the infrared image and the visible light image, respectively.

[0033] Optionally, the feature alignment unit performs feature alignment on the infrared image and the visible light image, including:

[0034] A 3×3 convolution kernel with a step size of 2 is used to downsample the basic features corresponding to the infrared and visible light images to obtain corresponding feature maps with a size halved; and the number of channels of the feature map corresponding to the visible light image is expanded to 12, and the number of channels of the feature map corresponding to the infrared image is expanded to 4.

[0035] Optionally, the multi-level feature enhancement focusing unit includes: a multi-scale field of view expansion convolution module, a first channel attention module and a first spatial pixel gain module.

[0036] Optionally, the multi-scale field of view expansion convolution module extracts local features of the visible light image and the infrared image at different scales by using 3×3 dilated convolution with dilation factors of 1, 2, 3, and 4 to obtain a multi-level feature map of the visible light image and a multi-level feature map of the infrared image.

[0037] Optionally, the calculation formula corresponding to the multi-level feature map of the visible light image is expressed as:

[0038] ;

[0039] ;

[0040] In the formula, Represents the feature maps extracted by convolution kernels of different scales, , represents the use of 3×3 dilated convolutional layers with dilation parameters 1, 2, 3, and 4, Represents the basic features of the visible light image obtained after the downsampling operation, Indicates the number of images processed in each batch, represents the number of channels of the feature map, Indicates height, Indicates length; , represents the multi-level feature map of the visible light image, Represents the concatenated fusion process of feature maps.

[0041] Optionally, the number of channels of the multi-level feature map of the visible light image and the multi-level feature map of the infrared image are both set to 24.

[0042] Optionally, the first channel attention module obtains a channel weighted feature map of the visible light image and a channel weighted feature map of the infrared image by assigning different weights to channels corresponding to the multi-level feature map of the visible light image and the multi-level feature map of the infrared image.

[0043] Optionally, different weights are assigned to channels corresponding to the multi-level feature map of the visible light image to obtain a channel weighted feature map of the visible light image, including:

[0044] The feature maps extracted by convolution kernels of different scales are calculated to obtain their respective channel attention weights, which are expressed as:

[0045] ;

[0046] In the formula, represents the channel attention weight, , Indicates application Attention mechanism.

[0047] Optionally, assigning different weights to channels corresponding to the multi-level feature map of the visible light image to obtain a channel weighted feature map of the visible light image further includes:

[0048] use The channel attention weights of each function are normalized and corrected to obtain The normalized and corrected attention weights of the channels are expressed as:

[0049] ;

[0050] In the formula, Indicates The first The attention weight of each channel, express The actual execution of the function, Indicates A raw value or measurement related to a channel, Indicates the 4 channels from 0 to 3 Perform the summation.

[0051] Optionally, assigning different weights to channels corresponding to the multi-level feature map of the visible light image to obtain a channel weighted feature map of the visible light image further includes:

[0052] The normalized and corrected attention weights of each channel are connected in series to obtain the attention weights of all channels corresponding to the feature information of the visible light image, which is expressed as:

[0053] ;

[0054] In the formula, Represents the attention weights of all channels of visible light feature information, Represents the concatenated fusion process of feature maps.

[0055] Optionally, assigning different weights to channels corresponding to the multi-level feature map of the visible light image to obtain a channel weighted feature map of the visible light image further includes:

[0056] Based on the multi-level feature map of the visible light image and all channel attention weights, the channel weighted feature map of the visible light image is obtained, which is expressed as:

[0057] ;

[0058] In the formula, Represents the channel weighted feature map of the visible light image, Represented as an element-wise multiplication operation between channels.

[0059] Optionally, the first spatial pixel gain module obtains a spatial pixel weighted feature map of the visible light image and a spatial pixel weighted feature map of the infrared image by assigning weights to pixels at each spatial position in the channel weighted feature map of the visible light image and the channel weighted feature map of the infrared image.

[0060] Optionally, assigning weights to pixels at each spatial position in the channel weighted feature map of the visible light image to obtain a spatial pixel weighted feature map of the visible light image includes:

[0061] Perform average pooling and maximum pooling operations on the channel weighted feature map of the visible light image to generate and , expressed as:

[0062] ;

[0063] In the formula, , represents the channel weighted feature map of the visible light image after average pooling, , represents the channel weighted feature map of the visible light image after maximum pooling, represents average pooling processing, Represents the maximum pooling process.

[0064] Optionally, assigning weights to pixels at each spatial position in the channel weighted feature map of the visible light image to obtain a spatial pixel weighted feature map of the visible light image further includes:

[0065] The corresponding two pooling results are spliced ​​to obtain , expressed as:

[0066] ;

[0067] In the formula, , represents the channel weighted feature map of the visible light image after splicing.

[0068] Optionally, assigning weights to pixels at each spatial position in the channel weighted feature map of the visible light image to obtain a spatial pixel weighted feature map of the visible light image further includes:

[0069] The number of channels is reduced to 1 through a 7×7 convolutional layer, and the The function performs activation processing and obtains the spatial pixel attention weight, which is expressed as:

[0070] ;

[0071] In the formula, , represents the spatial pixel attention weight, Represents a 7×7 convolution operation.

[0072] Optionally, assigning weights to pixels at each spatial position in the channel weighted feature map of the visible light image to obtain a spatial pixel weighted feature map of the visible light image further includes:

[0073] The spatial pixel attention weight is applied to the channel weighted feature map of the visible light image to obtain the spatial pixel weighted feature map of the visible light image, which is expressed as:

[0074] ;

[0075] In the formula, , represents the spatial pixel weighted feature map of the visible light image, Represents a pixel-by-pixel multiplication operation.

[0076] Optionally, the infrared and visible light feature fusion unit includes: a 1×1 convolution layer, a second spatial pixel gain unit and a second channel attention module, which is used to obtain the comprehensive features of the fused infrared and visible light images through fusion processing, and import the comprehensive features of the fused infrared and visible light images into the deep structure Layer2 to Layer5 stage of the YOLOv5s network to complete the secondary detection of active targets.

[0077] Optionally, the extracted spatial pixel weighted feature map of the visible light image and the spatial pixel weighted feature map of the infrared image are concatenated, and a 1×1 convolution kernel is used to perform a convolution fusion operation on each pixel, which is expressed as: ;

[0078] In the formula, , represents the fused infrared and visible light feature variables, Represents the spatial pixel weighted feature map of the visible light image, Represents the spatial pixel weighted feature map of the infrared image, represents a 1×1 convolution operation, Indicates a splicing action.

[0079] Optionally, different weights are assigned to pixels at different positions in the fused infrared and visible light feature variables through a second spatial pixel attention module, which is expressed as:

[0080] ;

[0081] In the formula, , represents the feature map generated after processing by the spatial pixel attention mechanism, express Application of activation function, represents the batch normalization process, Represents the extraction of spatial pixel attention distribution from the feature map.

[0082] Optionally, a second channel attention module is used to perform weighted integration on the channel dimension on the feature map generated after the spatial pixel attention mechanism is processed to obtain the comprehensive features of the fused infrared and visible light images, which can be expressed as: ;

[0083] In the formula, , represents the comprehensive features of the fused infrared and visible light images, Represents the channel attention weight calculation of the feature map.

[0084] Compared with the prior art, the present invention has the following beneficial effects:

[0085] (1) The present invention not only improves the original target detection network architecture of YOLOv5s, so that the weakening of target feature information in complex and harsh environments and the reduction of target size caused by long-distance imaging are effectively solved, and all-round and high-precision detection of targets in security areas is achieved; at the same time, a variety of non-visual sensors are integrated to achieve comprehensive coverage of targets at different distances, avoiding the misjudgment of a single sensor and reducing monitoring blind spots. The combination of multiple sensors not only consumes less power, has a shorter lifespan, and has a smaller amount of calculation than visual sensors, but also enables the system to adapt to the needs of different environments and complex scenes, and has strong practicality;

[0086] (2) The present invention also uses a layered detection strategy to pre-detect targets in the boundary area through multiple sensors to provide timely information to the system. Once a target is detected, the system can quickly activate the infrared and visible light fusion camera to capture and record images at a deeper level to ensure that security incidents are handled in a timely and effective manner. Not only is the energy consumption of monitoring equipment reduced, but the efficiency and accuracy of security monitoring are also significantly improved;

[0087] (3) The present invention also adds a remote hot backup unit to ensure data integrity and business continuity when a system failure occurs. This backup and recovery mechanism greatly reduces the risk of data loss and provides a strong guarantee for the stable operation of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 The figure shows the overall structure diagram of the monitoring mechanism in the regional security system integrating target detection of the present invention;

[0089] Figure 2It is a schematic diagram of the structure of the first monitoring mechanism in the regional security system integrating target detection of the present invention;

[0090] Figure 3 It is a schematic diagram of the networking of the control unit and the monitoring mechanism in the regional security system integrating target detection of the present invention;

[0091] Figure 4 The figure shows a schematic diagram of the target detection process of infrared and visible light images in the regional security method of fusion target detection of the present invention;

[0092] Figure 5 It is a schematic diagram of the framework of a multi-level feature enhancement focusing unit in the regional security method of fusion target detection of the present invention;

[0093] Figure 6 It is a schematic diagram of the framework of the infrared and visible light feature fusion unit in the regional security method of fusion target detection of the present invention;

[0094] Figure 7 It is a schematic diagram of a confusion matrix of infrared and visible light fusion target detection in the regional security method of fusion target detection of the present invention;

[0095] In the figure: 1-control unit; 2-fixing rod; 3-first monitoring mechanism; 4-second monitoring mechanism; 5-infrared and visible light cameras; 6-infrared sensor; 7-ultrasonic sensor; 8-long-range millimeter-wave radar; 9-short-range millimeter-wave radar; 10-single-spherical lightning rod antenna; 11-sound and light alarm; 12-wind power generation module; 13-solar power generation module; 14-fixing threaded hole; 15-reinforcement column; 16-disc chassis; 17-barrier wall; 18 control box. DETAILED DESCRIPTION

[0096] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0097] Example 1

[0098] refer to Figure 1 and Figure 3 , this embodiment introduces a regional security system integrating target detection, which includes: a control unit 1, and a plurality of monitoring mechanisms arranged on one side of the security area, each monitoring mechanism forming a boundary of the security area, and communicating with the control unit 1 in a wireless or wired manner;

[0099] The monitoring mechanism comprises: a fixed rod 2, and a first monitoring mechanism 3 and a second monitoring mechanism 4 arranged at the middle and top of the fixed rod 2;

[0100] The first monitoring mechanism 3 includes a plurality of sensors for pre-detecting moving targets and transmitting first target reconnaissance information corresponding to the plurality of sensors detected to the control unit 1;

[0101] The second monitoring mechanism 4 includes: an infrared and visible light camera 5 and an image processing unit, which are used to respond to the activation instruction output by the control unit 1, perform secondary detection on the active target, and transmit the corresponding second target reconnaissance information to the control unit 1; wherein the infrared and visible light camera 5 can not only take clear pictures through visible light during the day or under conditions of sufficient light, but also take effective images through infrared light at night or in a dark environment, and effectively detect the corresponding target image by forming an image fused by infrared and visible light; wherein the infrared and visible light camera is a spherical fusion camera;

[0102] The control unit 1 is respectively connected to the first monitoring mechanism 3 and the second monitoring mechanism 4 for communication, and is used to determine whether it is necessary to activate the infrared and visible light cameras 5 for secondary detection according to the first target reconnaissance information and the corresponding preset rules, and if so, control the execution of the preset security strategy according to the first target reconnaissance information and the second target reconnaissance information obtained in sequence;

[0103] The image processing unit performs feature alignment, extraction and fusion processing on the infrared images and visible light images collected by the infrared and visible light cameras through an improved YOLOv5s target detection network.

[0104] In specific applications of this embodiment, the sensor of the first monitoring mechanism 3 performs preliminary detection on active targets in the security area, and transmits the detected first target reconnaissance information to the control unit 1. The control unit 1 determines whether it is necessary to activate the second monitoring mechanism 4 for secondary detection based on the received first target reconnaissance information and corresponding preset rules. If it is determined to be necessary, an activation instruction is issued to the second monitoring mechanism 4. The infrared and visible light cameras 5 of the second monitoring mechanism 4 perform secondary detection on the target and transmit the collected image to the image processing unit. The image processing unit uses the improved YOLOv5s target detection network to perform feature alignment, extraction and fusion processing on the image to obtain the second target reconnaissance information. Finally, the control unit executes the preset security strategy, such as alarm, recording, tracking, etc., based on the first target reconnaissance information and the second target reconnaissance information obtained in sequence.

[0105] Example 2

[0106] On the basis of Example 1, Figures 1 to 3 As shown, this embodiment also has the following design:

[0107] The first monitoring mechanism 3 includes: an infrared sensor 6 , a long-range millimeter-wave radar 8 , a short-range millimeter-wave radar 9 , and an ultrasonic sensor 7 . The signal output ends of the infrared sensor 6 , the millimeter-wave radar, and the ultrasonic sensor 7 are all connected to the control unit 1 .

[0108] Specifically, the sensing range of infrared sensors generally ranges from a few centimeters to tens of meters. Infrared rays are used for data processing to achieve contactless temperature measurement and detection. It has the advantages of high sensitivity, fast response, and no need for direct contact with the object being measured.

[0109] As a long-distance detection device, the long-distance millimeter-wave radar 8 has a long detection range, which can usually cover a range of tens to hundreds of meters, and is not affected by weather conditions such as light, dust, haze, rain and snow, and can effectively detect suspicious targets at a long distance. As a short-distance detection device, the short-distance millimeter-wave radar 9 has a measurement range of only a few meters to tens of meters, and is mainly used for short-distance target detection, ranging, speed measurement, etc. However, based on its high-precision ranging function, the short-distance millimeter-wave radar 9 can accurately detect suspicious targets close to the border, providing an effective decision-making basis for safety warnings.

[0110] The ultrasonic sensor 7 is mainly used for distance measurement, speed measurement, non-contact detection of target surface position, and measurement of changing target size. In a regional security system, the ultrasonic sensor 7 can be used to accurately measure the distance of objects close to the boundary to provide support for preventing accidental intrusion or other security measures.

[0111] During the monitoring process, when the first monitoring mechanism (including infrared sensor, millimeter wave radar, ultrasonic sensor) detects a moving target, it will determine whether to activate the infrared and visible light cameras for secondary detection according to the following preset rules, as described below:

[0112] If the target moves faster than 3 meters per second (equivalent to the speed of a person running fast or a vehicle driving at a low speed), the camera will be activated immediately; the millimeter wave radar and ultrasonic sensors will jointly evaluate the shape and size of the target. If the match degree with the preset "human" or "vehicle" model is higher than 80%, and the target size is between 0.5 square meters and 50 square meters, the camera will be triggered to start; if the target stays in a sensitive area for more than 60 seconds, the system will automatically regard it as abnormal behavior and activate the camera for further observation; considering the importance of the motion trajectory, if the target's movement trajectory deviates from the preset safe path by more than 10 meters, or crosses the preset virtual boundary twice in a row, the camera will be activated and the event will be recorded.

[0113] It is worth noting that in order to improve the accuracy of judgment, the system will integrate the data of multiple sensors and require at least two sensors (such as infrared sensors and millimeter-wave radars) to simultaneously meet one of the above conditions before the camera can be activated. Secondly, in view of the time factor, especially at night or in low light conditions, the system will automatically adjust the movement speed threshold to 1 meter per second and reduce the shape size matching requirement to 90% to adapt to the challenges brought by environmental changes. In addition, the system will also consider the impact of environmental factors such as weather and light on sensor performance, and dynamically adjust the judgment logic through the built-in algorithm to ensure that security tasks can be accurately and reliably performed under various environmental conditions.

[0114] The monitoring mechanism further comprises: a single-spherical lightning rod antenna 10, an audible and visual alarm 11, and a power supply unit for providing power for monitoring operation.

[0115] Specifically, the single-spherical lightning rod antenna 10 is a lightning arrester with a large coverage area and a strong lightning discharge capability. The single-spherical lightning rod antenna 10 is installed at the top of the fixed pole, and is used to transmit the lightning current after the lightning arrestment to the ground through the internal metal inner conductor and the down conductor, thereby protecting the boundary device node from lightning side strikes and electrification. In most cases, it can ensure that the impact of the electrostatic field cable is less than 1 / 10 of the impedance of the boundary device node, avoiding the electrification of the boundary device node, so that the flashover phenomenon is effectively eliminated.

[0116] The sound and light alarm 11 is a sound and light alarm device, and its control end is connected to the control unit 1. When a suspicious person or other target appears on the scene, the control unit 1 controls the sound and light alarm 11 to work according to the first target reconnaissance information and the second target reconnaissance information. Specifically, it can send out a strong sound and light alarm signal to warn and drive away the on-site personnel to leave the boundary area, and prevent suspicious personnel from breaking into the security area. In addition, the sound and light alarm 11 can also form a sound and light alarm system with door magnets, infrared detectors and other equipment to remind the control center to pay attention to security threats such as illegal intrusion.

[0117] The power supply unit includes: a power generation module and a battery pack; the power generation module is installed at the top of the fixing rod 2, and its power output end is connected to the battery pack. It is worth noting that the power supply unit and the image processing unit are both arranged in the corresponding control box 18, and the control box 18 also includes a signal transceiver.

[0118] Specifically, the signal transceiver is a device in the control box 18 that is responsible for receiving and sending signals, that is, corresponding to the reception of signals and the sending of signals. The former corresponds to receiving signals from the control unit 1 or the remote hot backup unit, such as control instructions, data input, etc.; the latter corresponds to sending the signals or data inside the control box 18 to external devices or systems, such as the output of control commands, target detection results, and feedback of microcomputer processing information.

[0119] As a power supply device, the battery unit can provide stable power to components such as signal transceivers, microcomputers (image processing units), etc. to ensure their normal operation. Especially when the external power supply is interrupted or unstable, the battery can use its own stored energy or the clean reserve energy stored in the power generation module to maintain the basic operation of the control box and ensure the reliability and safety of the system.

[0120] The power generation module includes: a solar power generation module 13 that converts solar energy into electrical energy, and a wind power generation module 12 that converts wind energy into electrical energy. The electrical energy converted by the former can be used directly, or sent to the battery pack in the control box 18 for storage, or used to drive the load; the latter, driven by wind power, makes the windmill blades rotate, thereby driving the generator to generate electricity. The above two power generation methods not only provide clean energy for the monitoring agency and reduce environmental pollution, but also effectively improve the efficiency of energy utilization, and are highly practical.

[0121] The microcomputer (image processing unit) is the core processing unit in the control box 18, which is mainly responsible for data processing and the execution of control commands. Related functions include: data processing, control decision-making, communication management and system monitoring. Among them, data processing: processing the image screen, video frame and target detection taken by the second monitoring mechanism 4, receiving the signal or data from the signal transceiver, and performing necessary processing and analysis, such as conversion, storage, etc.; control decision-making: making control decisions based on the processed data or the preset control algorithm, such as the output of control commands, adjustment of system status, etc.; communication management: communicating with external devices or systems through the signal transceiver to realize data exchange and transmission of control instructions; system monitoring: monitoring the operating status of the entire control box, including the working conditions of each component, power status, etc., and taking corresponding measures when necessary to ensure the stable operation of the system.

[0122] In addition, the second monitoring mechanism 4 includes an illumination lamp, which can fill in the boundary area with light when the light is weak or the target is blurred, so as to make the target image clearer.

[0123] The regional security system based on fusion target detection also includes an off-site hot backup unit, which is located in other areas far away from the security area. It communicates with the control unit 1, synchronizes the data of the control unit and provides data backup and disaster recovery capabilities. It complements the control unit in terms of functions to ensure data integrity and business continuity. The specific functions and effects are described as follows.

[0124] Data backup, improve data security: The remote hot backup center backs up the data of the main system of regional security to the backup site in real time through the network in a synchronous or asynchronous manner. This backup method can reduce the risk caused by the failure of a single storage device or server, so that even if the main system fails, the backup center can quickly provide a complete copy of the data to ensure the normal operation of the business; Business continuity guarantee, enhance business availability: The backup site not only backs up data, but also can undertake certain business processing functions. When a disaster occurs, the backup center can quickly take over the business of the main system to ensure the continuity and availability of security services; Optimize resource utilization: The backup site usually only backs up data and does not undertake all business. However, if conditions permit, the backup center can also be used as an additional computing resource to handle other tasks to improve the overall resource utilization efficiency; Disaster recovery capability: In the face of severe situations such as natural disasters and man-made damage, the remote hot backup center can provide powerful disaster recovery capabilities. By quickly recovering data and taking over business, it ensures that the regional security system resumes normal operation in the shortest time and reduces the security risks caused by system failures.

[0125] The control unit 1 is located inside the security area and is responsible for monitoring and controlling the security system of the entire area to ensure the normal operation of the system. The control unit 1 includes: a master control computer, a signal transceiver, a backup power supply and an intercom.

[0126] Specifically, the master control computer is the core brain of the regional security system, responsible for the management and control of the entire system. It receives information from various security devices, such as video surveillance, intrusion alarm, access control management, etc., and performs centralized processing and analysis. Through the master control computer, the working status of each security device can be monitored in real time, and abnormal situations can be discovered and handled in a timely manner. In addition, the master control computer can automatically execute preset security strategies, such as automatic alarm, automatic recording, etc., to improve the automation level and processing efficiency of the system; in addition, by storing data such as video recording and alarm records in real time, it can facilitate the subsequent rapid retrieval of relevant information and further investigation and evidence collection.

[0127] As a bridge for communication between various devices in the security system, the transceiver is responsible for accurately transmitting various monitoring images and alarm signals to the control center, effectively realizing data exchange and information sharing between devices. Correspondingly, the control center can monitor and control remote security equipment to improve the coverage and flexibility of the system.

[0128] When the mains power outage or power failure occurs, the backup power supply can automatically switch to power the security system to ensure that the system can continue to operate stably. Secondly, the backup power supply can solve the problem of drastic fluctuations in network voltage and provide a stable power supply environment for the security system. In addition, the backup power supply is usually equipped with an intelligent management system, which monitors the corresponding power status and power level and other information in real time to facilitate management personnel to carry out timely and effective maintenance and replacement.

[0129] Intercom is an important tool for real-time voice communication between the control center and on-site personnel. It can improve the speed and accuracy of information transmission, and facilitate on-site personnel to report situations and receive instructions in a timely manner. In addition, intercom can also be integrated with other security equipment to respond and process the alarm information of the equipment in a timely manner to protect the safety of the area.

[0130] In addition to the fixing column 2, the present embodiment is also designed with a reinforcement column 15, a disc-shaped base 16 and fixing bolts, as described in detail as follows:

[0131] When subjected to forces from above or from the side, the lower fixing rod 2 can effectively transfer these loads to the reinforcement column 15 below and the ground to ensure the stability of the entire structure; in addition, by designing a corresponding adjustment mechanism to adjust the length and angle of the lower fixing rod 2, the height and direction of the entire fixing device can be adjusted to meet different installation requirements.

[0132] Specifically, the lower section of the fixed rod 2 is a multi-stage telescopic structure, including: an inner rod and an outer cylinder, the outer cylinder is provided with a plurality of positioning holes arranged equidistantly in the vertical direction, the inner rod can slide along the slide groove inside the outer cylinder, and is locked into the positioning hole on the outer cylinder through the corresponding locking bolt, so as to achieve precise adjustment of the length. In addition, in order to enhance stability, anti-slip patterns or rubber washers can be provided on the contact surface of the inner and outer rods to prevent sliding when subjected to force.

[0133] The reinforcement column 15 is arranged between the lower fixing rod 2 and the disc-shaped base 16. By increasing the vertical supporting force, the stability of the entire structure is improved to prevent shaking or tilting due to external forces. In addition, the reinforcement column 15 disperses the load transmitted from the lower fixing rod 2 to the ground, so that the pressure per unit area of ​​the ground is reduced, thereby enhancing the durability of the structure. In severe weather conditions, the reinforcement column 15 can effectively reduce the impact of external forces on the entire structure and ensure the safety of the fixing device.

[0134] The disc-shaped base 16 effectively disperses the load from above and reduces the pressure on the ground by increasing the contact area with the ground; in addition, by adding corresponding anti-slip textures or mounting holes on the disc-shaped base 16, a firm connection between the fixing device and the ground can be ensured to prevent safety accidents caused by sliding or movement.

[0135] The disc-shaped base 16 is correspondingly provided with fixing threaded holes 14 adapted to the fixing bolts. The fixing bolts are important components for connecting the lower fixing rod 2, the reinforcement column 15 and the disc-shaped base 16. The three are tightly connected together by tightening to form a stable overall structure. At the same time, by adjusting the fixing bolts, the height, direction and angle of the entire fixing device can be fine-tuned to achieve the best installation effect. The fixing bolts adopt anti-loosening designs, such as thread locking, elastic gaskets, etc., to ensure that they will not loosen or fall off during long-term use.

[0136] Example 3

[0137] This embodiment provides a regional security method for fusion target detection applicable to the regional security system for fusion target detection described in Embodiment 1 or Embodiment 2, which includes:

[0138] Acquire first target reconnaissance information obtained by the first monitoring agency through pre-detection;

[0139] According to the first target reconnaissance information and the corresponding preset rules, determine whether it is necessary to activate the second monitoring mechanism for secondary detection; if so, obtain the second target reconnaissance information obtained by the second monitoring mechanism after the secondary detection;

[0140] Based on the first target reconnaissance information and the second target reconnaissance information, completing the execution of the preset security strategy;

[0141] The processing of the second target reconnaissance information includes:

[0142] The infrared image and visible light image collected by the second monitoring mechanism are input into the improved YOLOv5s target detection network for feature alignment, extraction and fusion processing to obtain the second target reconnaissance information.

[0143] In the specific application of this embodiment, first, the sensor equipment is used for preliminary monitoring, video or image data is captured and processed, and preliminary target reconnaissance information is extracted. Based on the information provided by the first monitoring agency, it is determined whether further confirmation or detailed detection is needed through preset rules to avoid unnecessary waste of resources, that is, only when a suspected important event or target is found, a higher-precision monitoring device is started. The corresponding second monitoring agency usually has a higher resolution, a longer monitoring distance or a special spectral perception ability, and can provide more detailed and accurate image data; further, the control unit will decide what response measures to take based on the first and second target reconnaissance information, combined with preset security rules or strategies, which may include alarm notification, video recording storage, automatic tracking, and even linkage with on-site security personnel. In short, the method provided by this embodiment realizes efficient management of regional security through multi-level monitoring and detection, combined with advanced image processing and target detection technology, and effectively improves the intelligence level of the security system.

[0144] Example 4

[0145] Based on Example 3, this example also has the following design.

[0146] The infrared and visible light cameras are equipped with an improved YOLOv5s target detection network to solve two main problems: the weakening of target feature information in complex and harsh environments, and the reduction of target size caused by long-distance imaging, as described below.

[0147] Regarding question 1: The regional security system must be able to adapt to various complex environments such as dark nights, rain and fog and other bad weather and achieve all-weather reconnaissance and alert capabilities. However, the current mainstream visible light and infrared reconnaissance equipment is limited by the principles of physical imaging. Under the influence of various factors such as uneven lighting, temperature differences, and rain and fog interference, the contrast of the target in the image is reduced, and the visual features are blurred, resulting in weak feature intensity, which greatly increases the difficulty of target detection. Specifically, at night in backlight or weak light environments, the target outline in the visible light image is blurred and visibility is low, which brings great trouble to detection; although infrared images are not restricted by lighting conditions, when the temperature is similar, the boundary between the target and the background is blurred and it is easy to be submerged by the background. In addition, its lack of color information and weak internal texture details further increase the detection challenge.

[0148] Therefore, the infrared and visible light cameras equipped with the improved YOLOv5s target detection network mainly focus on effectively overcoming the suppression of target feature information by external adverse factors, strengthening the weakened target feature expression, and improving the detection accuracy of weak targets.

[0149] Regarding the second question: In regional security, in order to detect suspicious targets in advance and increase the warning time, reconnaissance imaging equipment needs to continuously extend the shooting distance. In long-distance imaging scenarios, the size of targets such as people and vehicles in the image is greatly reduced, the number of pixels is limited, most of the image area is invalid background, and the effective target area accounts for a very low proportion, which makes target feature extraction extremely difficult. In addition, deep neural networks usually use multiple downsampling to obtain deep semantic features in the process of extracting image features, which will result in a certain amount of information loss. For small-scale targets, this information loss is particularly obvious, because they originally occupy a very small proportion in the image. After multiple downsampling, the originally scarce effective information will be further reduced, which may lead to inaccurate detection due to the loss of a large amount of detail information.

[0150] Therefore, the infrared and visible light cameras are equipped with an improved YOLOv5s target detection network, which is also designed to more effectively discover and utilize the feature information of small targets.

[0151] In addition, infrared and visible light cameras also have the functions of video monitoring and temperature measurement. Specifically, the camera can monitor the scene in real time, display the details of the scene through visible light imaging, and combine infrared imaging to maintain monitoring capabilities at night or in low light; for key areas, the fusion camera can detect the temperature, feedback the temperature information in real time, and trigger other devices to link when the temperature exceeds the set threshold; in terms of target recognition and detection, based on the human pixel grid comparison, the spherical fusion camera module can automatically identify the personnel in the monitoring area, and compared with traditional visible light recognition, it is less affected by changes in ambient light and has a higher recognition accuracy; the system can automatically detect the behavior of people or vehicles entering or leaving the security area, which is suitable for regional protection.

[0152] It is worth noting that the current image fusion detection technology using deep learning technology can be roughly divided into initial fusion, mid-term fusion and late fusion strategies according to the differences in the fusion stage. In the initial fusion stage, pixel-level mixed assimilation is usually used, or deep fusion of the feature level is started at the beginning of the neural network; in the transition to the mid-term fusion stage, a comprehensive fusion operation will be carried out on the received intermediate layer features during the feature extraction process of the backbone network; as for the late fusion, it tends to perform feature-level or decision-level fusion in the later stages of the network.

[0153] The mid-term and late fusion methods are conducive to more comprehensive extraction and splicing of the feature information of the dual-modal image, but because they use a parallel dual-path network architecture, the parameter scale is huge and the computational load is increased, which makes the detection speed relatively slow; although the pixel-level fusion method in the initial fusion largely avoids increasing the number of parameters and computational burden of the detection network, it cannot fully utilize all the features of the dual-modal image. However, the use of feature fusion technology in the initial stage can effectively extract the infrared and visible light fusion features while avoiding a significant increase in the number of parameters and computational complexity.

[0154] In view of this, the improved YOLOv5s target detection network reconstructs the original network architecture of YOLOv5s, realizes the deep fusion of infrared and visible light image features at an early stage, and studies a novel method based on this to perform effective cross-modal detection of small targets.

[0155] Specifically, the improved YOLOv5s target detection network includes: a feature alignment unit, a multi-level feature enhancement focusing unit, and an infrared and visible light feature fusion unit, which are used to align, extract, and fuse features of the infrared image and the visible light image, respectively. The specific functions of each unit are as follows:

[0156] (1) Feature alignment unit

[0157] refer to Figure 4 In the improved network structure, the feature alignment unit uses convolution operations of the same size and length to extract basic feature contents such as intrinsic texture attributes, edge features and color elements of infrared and visible light images, and realizes precise alignment of the two modal images at the feature level, laying a solid foundation for the in-depth extraction and fusion of dual-modal features in the next stage.

[0158] To ensure the consistency of infrared and visible light image features, the infrared and visible light images are downsampled by applying a 3×3 convolution kernel with a step size of 2, so that the size of the feature map is halved. In order to reduce the size while maintaining information integrity, the number of channels of the feature map is expanded to four times the original, that is, the original 3-channel visible light image is expanded to 12 channels, and the single-channel infrared image is expanded to 4 channels.

[0159] (2) Multi-level feature enhancement focusing unit

[0160] Reference 4 and Figure 5 The multi-level feature enhancement focusing unit MTFEM uses convolution kernels of different sizes to respectively explore the local features of infrared and visible light images at different scales, and combines the attention mechanism of channel dimension and spatial pixel dimension to specifically enhance useful information while suppressing irrelevant noise components.

[0161] The multi-level feature enhancement focusing unit MTFEM includes: a multi-scale field of view expansion convolution module, a first channel attention module, and a first spatial pixel gain module (Spatial Pixel Accentuation, SPA).

[0162] Specifically, MTFEM uses 3×3 dilated convolutions with dilation parameters of 2, 3, and 4 to replace the standard convolution operations of 5×5, 7×7, and 9×9. This effectively reduces the number of parameters and computational intensity while maintaining the original receptive field. Secondly, MTFEM abandons the use of grouped convolutions to eliminate the problem of reduced information interaction that may be caused by this. Finally, in addition to continuing the channel attention mechanism, MTFEM also adds a spatial pixel attention mechanism to better serve the target detection task, especially for the detection of small targets. It can more accurately lock and strengthen the spatial position information of the target, thereby effectively extracting key pixel features.

[0163] (2.1) Multi-scale field of view expansion convolution module

[0164] The multi-scale field of view expansion convolution module extracts the local features of the visible light image and the infrared image at different scales by using 3×3 dilated convolution with dilation factors of 1, 2, 3, and 4, and obtains a multi-level feature map of the visible light image and a multi-level feature map of the infrared image. It is worth noting that the effective field of view of these dilated convolutions is consistent with the standard 3×3, 5×5, 7×7, and 9×9 convolutions, but the parameter quantity and computational load are maintained at the same level as the simple 3×3 convolution, thereby significantly reducing the overall parameter and computational requirements.

[0165] Given that the texture details of infrared images are relatively weak, the use of dilated convolution can also reduce the over-reliance of convolutional neural networks on target texture features to some extent, thereby being more conducive to extracting non-texture features such as the overall shape and contour of infrared targets. In the multi-scale field of view expansion convolution component, a network architecture with a consistent structure is used to extract features from visible light and infrared images respectively. Here, the feature extraction process of visible light images is used as an example to illustrate. For visible light images, the mathematical expression of multi-scale local feature extraction is as follows:

[0166] ;

[0167] In the formula, , , represents the feature map extracted by convolution kernels of different scales, represents the use of 3×3 dilated convolutional layers with dilation parameters 1, 2, 3, and 4, Represents the basic features of the visible light image obtained after the downsampling operation, Indicates the number of images processed in each batch, represents the number of channels of the feature map, Indicates height, Indicates length, Represents an image.

[0168] Represents the number of channels of a baseline or original feature map. This value can be arbitrary and depends on the design of the network and the input of the network layer. is one-fourth of the number of reference channels. Therefore, the expression It covers various feature maps generated by convolution operations with different dilation rates. By combining these feature map sets, we can construct a complete set of visible light image feature maps covering multiple scales. The specific combination is shown as follows:

[0169] ;

[0170] In the formula, , represents the multi-level feature map of the visible light image, Represents the concatenated fusion process of feature maps.

[0171] Similarly, infrared images can also be represented by corresponding multi-level features, denoted as In this process, in order to achieve a balanced integration of the feature information of visible light and infrared images and ensure that the two contribute equally in the fusion stage, the number of channels of the feature maps of both can be set to 24 units. In this way, in the process of target detection in complex environments, the feature contribution of both visible light and infrared spectrum bands is exactly 50%, which effectively enhances the robustness of the overall detection performance.

[0172] (2.2) First channel attention module

[0173] As a classic channel attention mechanism, the channel attention module (Squeeze-and-Excitation) assigns different weights to different feature channels, focusing on amplifying effective channel information, thereby refining a more accurate target feature expression. First, the size of the feature map is reduced to 1×1 using the global average pooling technology; then, a fully connected layer is used to compress the number of channels to 1 / 16 of the original number, and the The activation function is used for processing; then, a fully connected layer is used to restore the number of channels to the original number, and then the The function normalizes the weights to obtain the corresponding attention weights for each channel.

[0174] For feature maps extracted from convolution kernels of different scales , first use The modules calculate their respective channel attention weights , and perform weighted processing on it; in order to establish the global correlation between features of different scales, the Function pair Perform normalization correction to generate ; Finally, by concatenating all the normalized correction results, we can obtain the attention weights of all channels of the visible light feature information .

[0175] That is to say, the first channel attention module can obtain the channel weighted feature map of the visible light image and the channel weighted feature map of the infrared image by assigning different weights to the channels corresponding to the multi-level feature map of the visible light image and the multi-level feature map of the infrared image.

[0176] Specifically, taking a visible light image as an example, different weights are assigned to channels corresponding to a multi-level feature map of the visible light image to obtain a channel weighted feature map of the visible light image, including the following steps:

[0177] First, the feature maps extracted by convolution kernels of different scales are calculated to obtain their respective channel attention weights, which are expressed as:

[0178] ;

[0179] In the formula, represents the channel attention weight, Indicates application Attention mechanism.

[0180] Further, using The channel attention weights of each function are normalized and corrected to obtain The normalized and corrected attention weights of the channels are expressed as: ;

[0181] In the formula, in the formula, Indicates The first The attention weight of each channel, express The actual execution of the function, Indicates A raw value or measurement related to a channel, Indicates the 4 channels from 0 to 3 Perform the summation.

[0182] Furthermore, the normalized and corrected attention weights of each channel are connected in series to obtain the attention weights of all channels corresponding to the feature information of the visible light image, which is expressed as:

[0183] ;

[0184] In the formula, Represents the attention weights of all channels of visible light feature information, Represents the concatenated fusion process of feature maps.

[0185] Furthermore, based on the multi-level feature map of the visible light image and all channel attention weights, a channel weighted feature map of the visible light image is obtained, which is expressed as:

[0186] ;

[0187] In the formula, Represents the channel weighted feature map of the visible light image, Represented as an element-wise multiplication operation between channels.

[0188] (2.3) First spatial pixel gain module

[0189] In order to further highlight the feature expression of weak targets and effectively capture the spatial position characteristics of targets, this embodiment also introduces a spatial pixel gain module based on the channel attention mechanism. The spatial pixel gain module is a process of assigning weights to pixels at each spatial position in an image by using the attention mechanism, so as to better focus on the spatial positioning information of the target.

[0190] First, the feature map Perform pixel value processing on the channel dimension, implement average pooling and maximum pooling operations respectively, and generate and ; Next, the two pooling results are concatenated to form ; Then, a 7×7 convolutional layer is used to reduce the number of channels to 1, and the Sigmoid function is used for activation to obtain the spatial pixel attention weights ; Finally, the spatial pixel attention weight Acting on feature maps , spatial pixel attention weights are assigned to the feature map to generate feature information optimized for visible light images.

[0191] That is to say, the first spatial pixel gain module can obtain the spatial pixel weighted feature map of the visible light image and the spatial pixel weighted feature map of the infrared image by assigning weights to each spatial position pixel in the channel weighted feature map of the visible light image and the channel weighted feature map of the infrared image.

[0192] Specifically, assigning weights to pixels at each spatial position in the channel weighted feature map of the visible light image to obtain a spatial pixel weighted feature map of the visible light image includes:

[0193] First, average pooling and maximum pooling operations are performed on the channel weighted feature map of the visible light image to generate and , expressed as:

[0194] ;

[0195] In the formula, , represents the channel weighted feature map of the visible light image after average pooling, , represents the channel weighted feature map of the visible light image after maximum pooling, represents average pooling processing, Represents the maximum pooling process.

[0196] Furthermore, the corresponding two pooling results are spliced ​​to obtain , expressed as:

[0197] ;

[0198] In the formula, , represents the channel weighted feature map of the visible light image after splicing.

[0199] Furthermore, the number of channels is reduced to 1 through a 7×7 convolutional layer, and the The function performs activation processing and obtains the spatial pixel attention weight, which is expressed as:

[0200] ;

[0201] In the formula, , represents the spatial pixel attention weight, Represents a 7×7 convolution operation.

[0202] Furthermore, the spatial pixel attention weight is applied to the channel weighted feature map of the visible light image to obtain a spatial pixel weighted feature map of the visible light image, which is expressed as:

[0203] ;

[0204] In the formula, , represents the spatial pixel weighted feature map of the visible light image, Represents the pixel-by-pixel multiplication operation. Through a similar processing flow, the feature information corresponding to the infrared image can also be obtained, which is recorded as .

[0205] (3) Infrared and visible light feature fusion unit

[0206] refer to Figure 4 and Figure 6 The infrared and visible light feature fusion unit uses the attention mechanism to gradually realize the adaptive fusion process of features, so as to efficiently explore and integrate the complementary feature information between the two modal images, thereby significantly improving the ability to represent weak target features. Finally, the fused comprehensive feature map will be sent to the Layer2 to Layer5 stage in the YOLOv5s backbone network to further extract high-level semantic features and perform target detection operations.

[0207] The infrared and visible light feature fusion unit includes: a 1×1 convolution layer, a second spatial pixel gain module and a second channel attention module, which aims to effectively integrate the advantageous feature information of the visible light and infrared images and weaken the influence of noise factors on the fusion results. The specific fusion process is as follows:

[0208] First, we process the feature information extracted from visible light and infrared images. and When , a splicing strategy is adopted to maximize the integration of the characteristics of the two modalities, and then a simple and efficient 1×1 convolution kernel is used to perform convolution fusion operations on each pixel to ensure that the number of channels remains constant after fusion. This process can be expressed by the following formula:

[0209] ;

[0210] In the formula, , represents the fused infrared and visible light feature variables, Represents the spatial pixel weighted feature map of the visible light image, Represents the spatial pixel weighted feature map of the infrared image, represents a 1×1 convolution operation, Indicates a splicing action.

[0211] Furthermore, based on the 1×1 convolution fusion feature, a spatial pixel attention module is further applied to assign different weights to pixels at different positions in the feature map. The use of spatial pixel attention technology can not only focus on subtle key feature signals more accurately and filter out unnecessary noise interference, but also help to more accurately distinguish foreground and background information for target detection tasks, thereby improving the positioning accuracy of target boundaries. This process can be abstracted as the following formula: ;

[0212] In the formula, , represents the feature map generated after processing by the spatial pixel attention mechanism, express Application of activation function, represents the batch normalization process, Represents the extraction of spatial pixel attention distribution from the feature map.

[0213] Furthermore, in order to enhance the intrinsic connection between fusion features in each channel and effectively improve the model's ability to understand and utilize global features, the infrared and visible light fusion unit adopts the SE attention mechanism, which performs weighted integration of features in the channel dimension to obtain the comprehensive features of the fused infrared and visible light images, aiming to capture the information contained in each channel. The corresponding mathematical expression is:

[0214] ;

[0215] In the formula, , represents the comprehensive features of the fused infrared and visible light images, Represents the channel attention weight calculation of the feature map.

[0216] It is worth noting that in the above fusion process, the 1×1 convolution weights, spatial pixel gain unit weights, and channel attention SE module weights used by the infrared and visible light feature fusion unit are all automatically learned and optimized through the stochastic gradient descent algorithm without the need for human intervention, thereby achieving the fusion of infrared and visible light features.

[0217] Finally, the comprehensive features of infrared and visible light images are fused It is imported into the deep structure Layer2 to Layer5 of the YOLOv5s network to further extract deep semantic features. Considering that shallow local detail features are crucial to constructing deep global semantic features, by extracting multi-scale local features of visible light and infrared images at the shallow level and integrating their complementary information, more detailed local feature information of the target can be obtained, which is extremely beneficial for the YOLO network to further refine the deep semantic features of the target. By constructing a feature pyramid structure and effectively fusing the shallow feature map with the deep feature map, the loss of detail information that may occur during the deep feature extraction process can be effectively compensated, thereby improving the accuracy and stability of weak target detection.

[0218] In addition, to further illustrate the effectiveness of the above improved target model, this embodiment is also verified through corresponding ablation experiments, as described below.

[0219] (1) Feature alignment unit ablation experiment

[0220] When selecting the convolution kernel size of the feature alignment unit, four different sizes of convolution kernels, 3×3, 5×5, 6×6, and 7×7, were compared and the results are shown in Table 1. The best results and suboptimal results are highlighted in bold and underlined. According to the experimental data, the 3×3 convolution kernel showed the best performance.

[0221] The reason is that although a larger convolution kernel can broaden the receptive field and help obtain a wider range of contextual information, in the process of finely extracting local features of an image, a large convolution kernel is often not conducive to capturing the detailed local features of the target, especially for the detection of tiny targets. In view of this, it was finally decided to use a 3×3 convolution kernel as the core component of the feature alignment unit.

[0222]

[0223] (2) Multi-level feature enhancement focus unit ablation experiment

[0224] In the construction of the MTFEM unit, the convolution kernel was ablated and three groups of conventional convolution kernel configurations with different sizes were tried, namely [3, 5, 7, 9], [1, 3, 5, 7] and [1, 3, 3, 5]. The performance of the dilated convolution with a size of [3, 3, 3, 3] and a dilation coefficient of [1, 2, 3, 4] was also investigated. The experimental results are shown in Table 2. In the table, the best and suboptimal results are highlighted in "bold and emphasis" and "underline" format, respectively.

[0225]

[0226] Experimental results show that when extracting effective features of infrared and visible light images, dilated convolution with different dilation coefficients performs well, effectively overcoming the problem of "over-reliance on texture" that may exist in conventional convolution, thus achieving the best performance in target detection.

[0227] In the original YOLOv5s model architecture, the number of output channels of Layer 1 is set to 32. and The number of channels was set to 16, 24, and 32 respectively, and ablation experiments were carried out. The experimental results are shown in Table 3. In the table, the best results and suboptimal results are marked by "bold and emphasis" and "underline" respectively.

[0228] The experimental results show that when and When the number of channels is 24, the model performs best. When the number of channels is reduced to 16, the target feature information is not sufficient to fully describe the target, resulting in a lower recall rate and an increased probability of missed detection. When the number of channels is increased to 32, the target feature information becomes more redundant, which not only fails to enhance the expression of effective features, but also interferes with its substantial information, thereby reducing the accuracy of the model. Therefore, the number of channels is finally set to 24 in this embodiment.

[0229]

[0230] (3) Ablation experiment of infrared and visible light feature fusion unit

[0231] In order to evaluate the improvement of the detection performance by the infrared and visible light feature fusion unit, this part is verified by a series of ablation experiments. The relevant results are shown in Table 4. Without the infrared and visible light feature fusion unit, the detection accuracy is 77.49%. With the gradual introduction of point-by-point convolution, channel attention module and spatial pixel gain unit SPAU, the detection accuracy is improved by 0.84%. This result shows that the proposed bimodal fusion module can effectively integrate bimodal features, thereby improving the accuracy of target detection.

[0232]

[0233] The results of the infrared and visible light fusion target detection method are shown in the confusion matrix diagram, where the horizontal axis represents the actual category and the vertical axis represents the category predicted by the model. Figure 7 In the confusion matrix shown, we can observe that:

[0234] For samples in the Car category, the model’s prediction accuracy is as high as 96.07%. This high accuracy is mainly due to the fact that the Drone Vehicle dataset has the most samples in the Car category, and the model has the highest degree of optimization of the loss function for the Car category during training, so it has the greatest advantage in identifying the Car category.

[0235] For the freight car samples, the prediction accuracy was 62.00%, of which the samples misidentified as trucks accounted for the highest proportion, reaching 19.80%. The prediction accuracy of the truck samples was 73.40%, and 10.65% of the samples were misidentified as freight cars, which also became the largest category of prediction errors in the Truck category. The reason may be that both freight cars and trucks have transportation functions and have certain similarities in appearance features, so they are easily confused during identification.

[0236] The prediction accuracy of samples in the Bus category is as high as 92.01%. Although the number of samples is relatively small, the false positive rate is low due to its distinct features and obvious differences from other categories. Finally, the prediction accuracy of Van samples is 59.89%, of which 27.66% are incorrectly predicted as Car. This is mainly because some sedan samples in the dataset are SUV models, and the appearance features of Van are very similar to those of SUV models, resulting in some Van samples being incorrectly marked as Car.

[0237] Example 5

[0238] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the regional security method integrating target detection as described in Embodiment 3 or Embodiment 4 are implemented.

[0239] It should be understood by those skilled in the art that the embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0240] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0241] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0242] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0243] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A regional security system based on fusion target detection, characterized in that: include: A control unit (1), and a plurality of monitoring mechanisms arranged on one side of a security area; the monitoring mechanisms include: a fixing rod (2), and a first monitoring mechanism (3) and a second monitoring mechanism (4) arranged at the middle and top of the fixing rod (2); The first monitoring mechanism (3) comprises a plurality of sensors for pre-detecting moving targets and transmitting first target reconnaissance information corresponding to the plurality of sensors detected to the control unit (1); The second monitoring mechanism (4) comprises: an infrared and visible light camera (5) and an image processing unit, which is used to respond to the activation instruction output by the control unit (1), perform secondary detection on the moving target, and transmit the corresponding second target reconnaissance information to the control unit (1); The control unit (1) is respectively connected to the first monitoring mechanism (3) and the second monitoring mechanism (4) for determining whether it is necessary to activate the infrared and visible light cameras (5) for secondary detection based on the first target detection information and corresponding preset rules, and if so, controlling the execution of the preset security strategy based on the first target detection information and the second target detection information obtained in sequence; The image processing unit performs feature alignment, extraction and fusion processing on the infrared image and visible light image collected by the infrared and visible light cameras (5) through an improved YOLOv5s target detection network; the improved YOLOv5s target detection network includes: a feature alignment unit, a multi-level feature enhancement focusing unit and an infrared and visible light feature fusion unit, which are respectively used to perform feature alignment, extraction and fusion processing on the infrared image and the visible light image; The feature alignment unit performs feature alignment on the infrared image and the visible light image, comprising: Down-sampling the basic features corresponding to the infrared and visible light images using a 3×3 convolution kernel with a step size of 2 to obtain a corresponding feature map with a size halved; and the number of channels of the feature map corresponding to the visible light image is expanded to 12, and the number of channels of the feature map corresponding to the infrared image is expanded to 4; The multi-level feature enhancement focusing unit includes: a multi-scale field of view expansion convolution module, a first channel attention module and a first spatial pixel gain module; The multi-scale field of view expansion convolution module extracts local features of the visible light image and the infrared image at different scales by using 3×3 dilated convolution with dilation factors of 1, 2, 3, and 4, thereby obtaining a multi-level feature map of the visible light image and a multi-level feature map of the infrared image; The calculation formula corresponding to the multi-level feature map of the visible light image is expressed as: ； ； In the formula, Represents the feature maps extracted by convolution kernels of different scales, , represents the use of 3×3 dilated convolutional layers with dilation parameters 1, 2, 3, and 4, Represents the basic features of the visible light image obtained after the downsampling operation, Indicates the number of images processed in each batch, represents the number of channels of the feature map, Indicates height, Indicates length; , represents the multi-level feature map of the visible light image, Represents the concatenated fusion process of feature maps; The number of channels of the multi-level feature map of the visible light image and the multi-level feature map of the infrared image are both set to 24.

2. The regional security system based on fusion target detection according to claim 1 is characterized in that: The first monitoring mechanism (3) comprises: an infrared sensor (6), a long-range millimeter wave radar (8), a short-range millimeter wave radar (9) and an ultrasonic sensor (7), and signal output ends of the infrared sensor (6), the long-range millimeter wave radar (8), the short-range millimeter wave radar (9) and the ultrasonic sensor (7) are all connected to the control unit (1); If at least two of the infrared sensor (6), the long-range millimeter-wave radar (8), the short-range millimeter-wave radar (9) and the ultrasonic sensor (7) simultaneously satisfy one of the preset rules, the infrared and visible light camera (5) is activated for secondary detection, and the preset rules include: The moving speed of the moving target exceeds 3 meters per second; the shape and size of the moving target match the preset "human" or "vehicle" model by more than 80%, and the size of the moving target is between 0.5 square meters and 50 square meters; the moving target stays in the preset sensitive area for more than 60 seconds; and the moving trajectory of the moving target deviates from the preset safe path by more than 10 meters or crosses the preset virtual boundary twice in succession; Among them, at night or under low light conditions, the moving speed is set to 1 meter per second, and the matching degree is set to 90%.

3. The regional security system based on fusion target detection according to claim 2 is characterized in that: The monitoring mechanism further comprises: a single-spherical lightning rod antenna (10), an audible and visual alarm (11), and a power supply unit for providing power for monitoring operation; The single-spherical lightning rod antenna (10) is installed at the top end of the fixing rod (2) and is used to transmit the lightning current after the lightning is connected to the ground through the internal metal inner conductor and the down conductor; The control end of the sound and light alarm (11) is connected to the control unit (1), and the control unit (1) controls the sound and light alarm (11) to operate according to the first target reconnaissance information and the second target reconnaissance information; The power supply unit comprises: a power generation module and a battery pack; the power generation module is mounted on the top end of the fixing rod (2), and its power output end is connected to the battery pack; The power generation module comprises: one or more of a wind power generation module (12) and a solar power generation module (13).

4. The regional security system based on fusion target detection according to claim 1 or 3, characterized in that: It also comprises a remote hot backup unit, which is in communication connection with the control unit (1) and is used for synchronizing relevant data stored in the control unit (1).

5. A regional security method based on fusion target detection, characterized in that: include: Acquiring first target reconnaissance information obtained by the first monitoring mechanism (3) through pre-detection; According to the first target reconnaissance information and corresponding preset rules, determining whether it is necessary to activate the second monitoring mechanism (4) for secondary detection; if so, obtaining the second target reconnaissance information obtained by the second monitoring mechanism (4) through secondary detection; Based on the first target reconnaissance information and the second target reconnaissance information, completing the execution of the preset security strategy; The processing of the second target reconnaissance information includes: The infrared image and the visible light image collected by the second monitoring mechanism (4) are input into an improved YOLOv5s target detection network for feature alignment, extraction and fusion processing, thereby obtaining second target reconnaissance information; the improved YOLOv5s target detection network comprises: a feature alignment unit, a multi-level feature enhancement focusing unit and an infrared and visible light feature fusion unit, which are respectively used to align, extract and fuse the features of the infrared image and the visible light image; The feature alignment unit performs feature alignment on the infrared image and the visible light image, comprising: Down-sampling the basic features corresponding to the infrared and visible light images using a 3×3 convolution kernel with a step size of 2 to obtain a corresponding feature map with a size halved; and the number of channels of the feature map corresponding to the visible light image is expanded to 12, and the number of channels of the feature map corresponding to the infrared image is expanded to 4; The multi-level feature enhancement focusing unit includes: a multi-scale field of view expansion convolution module, a first channel attention module and a first spatial pixel gain module; The multi-scale field of view expansion convolution module extracts local features of the visible light image and the infrared image at different scales by using 3×3 dilated convolution with dilation factors of 1, 2, 3, and 4, thereby obtaining a multi-level feature map of the visible light image and a multi-level feature map of the infrared image; The calculation formula corresponding to the multi-level feature map of the visible light image is expressed as: ； ； In the formula, Represents the feature maps extracted by convolution kernels of different scales, , represents the use of 3×3 dilated convolutional layers with dilation parameters 1, 2, 3, and 4, Represents the basic features of the visible light image obtained after the downsampling operation, Indicates the number of images processed in each batch, represents the number of channels of the feature map, Indicates height, Indicates length; , represents the multi-level feature map of the visible light image, Represents the concatenated fusion process of feature maps; The number of channels of the multi-level feature map of the visible light image and the multi-level feature map of the infrared image are both set to 24.

6. The regional security method based on fusion target detection according to claim 5 is characterized in that: The first channel attention module obtains a channel weighted feature map of the visible light image and a channel weighted feature map of the infrared image by assigning different weights to channels corresponding to the multi-level feature map of the visible light image and the multi-level feature map of the infrared image; Wherein, different weights are assigned to channels corresponding to the multi-level feature map of the visible light image to obtain a channel weighted feature map of the visible light image, including: The feature maps extracted by convolution kernels of different scales are calculated to obtain their respective channel attention weights, which are expressed as: ； In the formula, Indicates The first The attention weight of each channel, express The actual execution of the function, Indicates A raw value or measurement related to a channel, Indicates the 4 channels from 0 to 3 Perform summation; The normalized and corrected attention weights of each channel are connected in series to obtain the attention weights of all channels corresponding to the feature information of the visible light image, which is expressed as: ； In the formula, Represents the attention weights of all channels of visible light feature information, Represents the concatenated fusion process of feature maps; Based on the multi-level feature map of the visible light image and all channel attention weights, the channel weighted feature map of the visible light image is obtained, which is expressed as: ； In the formula, Represents the channel weighted feature map of the visible light image, Represented as an element-wise multiplication operation between channels.

7. The regional security method based on fusion target detection according to claim 6 is characterized in that: The first spatial pixel gain module obtains a spatial pixel weighted feature map of the visible light image and a spatial pixel weighted feature map of the infrared image by assigning weights to pixels at each spatial position in the channel weighted feature map of the visible light image and the channel weighted feature map of the infrared image; The step of assigning weights to pixels at each spatial position in the channel weighted feature map of the visible light image to obtain a spatial pixel weighted feature map of the visible light image includes: Perform average pooling and maximum pooling operations on the channel weighted feature map of the visible light image to generate and , expressed as: ； In the formula, , represents the channel weighted feature map of the visible light image after average pooling, , represents the channel weighted feature map of the visible light image after maximum pooling, represents average pooling processing, Represents the maximum pooling process; The corresponding two pooling results are spliced ​​to obtain , expressed as: ； In the formula, , represents the channel weighted feature map of the visible light image after splicing; The number of channels is reduced to 1 through a 7×7 convolutional layer, and the The function performs activation processing and obtains the spatial pixel attention weight, which is expressed as: ； In the formula, , represents the spatial pixel attention weight, Represents a 7×7 convolution operation; The spatial pixel attention weight is applied to the channel weighted feature map of the visible light image to obtain the spatial pixel weighted feature map of the visible light image, which is expressed as: ； In the formula, , represents the spatial pixel weighted feature map of the visible light image, Represents a pixel-by-pixel multiplication operation.

8. The regional security method based on fusion target detection according to claim 7 is characterized in that: The infrared and visible light feature fusion unit includes: a 1×1 convolution layer, a second spatial pixel gain unit and a second channel attention module, which is used to obtain the comprehensive features of the fused infrared and visible light images through fusion processing, and import the comprehensive features of the fused infrared and visible light images into the deep structure Layer2 to Layer5 stage of the YOLOv5s network to complete the secondary detection of the moving target; Among them, the spatial pixel weighted feature map of the extracted visible light image and the spatial pixel weighted feature map of the infrared image are spliced, and a 1×1 convolution kernel is used to perform a convolution fusion operation on each pixel, which is expressed as: ; In the formula, , represents the fused infrared and visible light feature variables, Represents the spatial pixel weighted feature map of the visible light image, Represents the spatial pixel weighted feature map of the infrared image, represents a 1×1 convolution operation, Indicates splicing action; Through the second spatial pixel attention module, different weights are assigned to pixels at different positions in the fused infrared and visible light feature variables, which can be expressed as: ； In the formula, , represents the feature map generated after processing by the spatial pixel attention mechanism, express Application of activation function, represents the batch normalization process, Represents the extraction of spatial pixel attention distribution from the feature map; Through the second channel attention module, the feature map generated after the spatial pixel attention mechanism is processed and weighted integration is performed on the channel dimension to obtain the comprehensive features of the fused infrared and visible light images, which can be expressed as: ; In the formula, , represents the comprehensive features of the fused infrared and visible light images, Represents the channel attention weight calculation of the feature map.

Images