Low-power-consumption intrusion detection and identification technology based on brain-like intelligence and storage medium

Abstract

The invention provides a low-power-consumption intrusion detection and identification technology based on brain-like intelligence, intrusion detection can be performed on pedestrians, vehicles and unmanned aerial vehicles invading a set area with low power consumption, and intrusion evidences are reserved. The scheme is suitable for the field of urban digital security and protection, can effectively improve the development of the field of urban security and protection, and provides lasting convenience for staff of an urban security and protection department. Through multiple experiments, the average power consumption of the dual-mode integrated intrusion detection equipment designed by the invention is below 2mW when no foreign object invades, the average power consumption is below 40mW when the foreign object invades, and a long-term stable intrusion detection task can be realized by arranging a battery.

Description

Technical Field

[0001] The present invention belongs to the technical field of intrusion detection and identification, and in particular proposes a low-power intrusion detection and identification technology based on brain-like intelligence. Background Art

[0002] With the continuous advancement of science and technology, the entire society is moving towards digitalization, and there is an urgent need for low-power detection equipment in the field of urban security. The widespread use of new sensing detection equipment such as various types of drones and tiny spacecraft, as well as the potential threat of surprise attacks in urban security, have made us face increasingly complex and extensive security challenges. The common characteristics of these challenges are that they involve a wide geographical range, extremely harsh on-site environments, and require real-time monitoring and rapid response. Therefore, long-lasting and extremely low-power detection equipment has become very urgent and important. These advanced devices can operate for a long time without causing unnecessary power consumption, while being able to provide high-precision monitoring data.

[0003] Dynamic event-driven visual sensors (DVS cameras) combined with brain-like SOC computing chips have extremely low power consumption and can achieve efficient and affordable monitoring without a large amount of power resources. This new sensor significantly reduces power consumption by generating events for data analysis only when the scene changes, rather than continuously capturing data. This allows the monitoring system to run longer without incurring high operation and maintenance costs. Monitoring systems based on DVS cameras can not only cope with the threats of various intruders, but also achieve a wide range of monitoring tasks in a more economical way. Therefore, event-driven visual sensors represent an innovative, low-cost, and energy-efficient monitoring method that can meet today's complex monitoring needs and bring new opportunities to the field of monitoring. Their role is not only to detect potential threat behaviors early, but also to support security forces to better understand threats so that they can take quick action.

[0004] Most existing technologies use visible light (infrared cameras) and auxiliary computing chips for intrusion detection, which includes the following process: First, a continuous video stream is captured from the monitored area and transmitted to the computing chip. On the computing chip, the video frames are preprocessed and feature extracted, and the CNN network detects and locates potential intrusion targets, such as people and vehicles. Subsequently, the system performs behavioral analysis to examine information such as the target's movement, trajectory, and interaction to determine whether there is an abnormal situation. If abnormal intrusion behavior is found, the system triggers an alarm and notifies relevant personnel to take measures. Finally, the system records and stores the video, analysis results, and alarm information for further review and legal tracking. This process fully utilizes the performance of computing chips and communication equipment, making intrusion detection more intelligent, efficient, and reliable. However, the RGB camera itself has high power consumption, and the operation of the convolutional neural network (CNN) also consumes a lot of electricity, which makes the above method face serious challenges in achieving low-power long-term operation. The continuous capture and transmission of high-resolution images of RGB cameras involves a lot of calculations and high energy consumption, which puts higher requirements on battery power or energy-saving power supply, and increases the maintenance cost of the equipment. Summary of the invention

[0005] One of the purposes of the present invention is to provide a low-power intrusion detection and identification technology based on brain-like intelligence, which can detect intrusion of pedestrians, vehicles, and drones that invade a set area with low power consumption and retain evidence of the intrusion.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] A low-power intrusion detection and identification technology based on brain-like intelligence, including:

[0008] Step 1: Data set processing and data processing enhancement;

[0009] Step 2: Build and deploy a CNN classification model based on SNN, and train a CNN binary classification model of the intrusion detection SNN architecture on the processed and enhanced data set to detect whether a target intrusion occurs in the set area;

[0010] Step 3: Intrusion detection and recognition and RGB device wake-up. Based on step 2, the target in the field of view of the DVS camera is identified. If the classification model recognizes a positive sample, a high-frequency pulse is emitted to wake up the RGB device for secondary detection verification. If the classification model recognizes a negative sample, it means that there is no target intrusion and detection continues.

[0011] Step 4: The RGB device identifies and retains the intrusion evidence. Based on steps 1 to 3, the RGB device captures the image and sends the captured image to the YOLO detection model. If the detection result is determined to be a positive sample of the target intrusion, the result is saved and a dial-up intrusion alarm is issued through the communication module. The RGB device is shut down and waits for the next wake-up. If the detection result is a negative sample, it is a false alarm, and the RGB device is shut down and waits for the next wake-up.

[0012] Preferably, the step 1 further comprises:

[0013] Step 1.1, data collection, including drone, vehicle, pedestrian and other category data sets, drone, vehicle, pedestrian three categories as positive samples, other category data sets including background, various other situations as negative samples, and the data set is split into training set, validation set and test set in the ratio of 8:1:1;

[0014] Step 1.2: Data preprocessing: converting the event information output by the DVS camera into image frame information that can be processed by the brain-like SOC chip model;

[0015] Step 1.3, data set enhancement, using one or more methods such as image inversion, brightness adjustment, random cropping, edge filling, saturation modification, mixup, etc. to enhance the data set; for the sorted training set, set a certain probability to randomly select a part of the data, add the above image enhancement method, and do not perform any data enhancement on the test set.

[0016] Preferably, the step 2 further comprises:

[0017] The frames after data preprocessing are further processed by the pulse convolutional neural network Dynap-CNN to obtain the category output of the current frame. The data frame will pass through the Conv module, IAFSqueeze module, Pool module, and FC (fully connected layer) in sequence to obtain the predicted category. The cross entropy loss is calculated between the predicted category and the true category, and the weights of the above model are optimized through back propagation, and it is continuously iterated until the model reaches the set accuracy.

[0018] Preferably, the Dynap-CNN includes 3 Conv modules, the number of convolution kernels of the first Conv module is 8, the size of the convolution kernel is 5×5, and the step size is 2; the number of convolution kernels of the second Conv module is 16, the size of the convolution kernel is 5×5, and the step size is 2; the number of convolution kernels of the third Conv module is 8, the size of the convolution kernel is 5×5, and the step size is 1.

[0019] Preferably, the step 3 further comprises:

[0020] Step 3.1, Model Post-Processing and Detection and Identification. The classification output of the Dynap-CNN model is a probability value P ranging from 0 to 1. By setting a threshold S, if P > S, it indicates that there is a potential target intrusion; if P < S, it indicates that there is no intrusion detection.

[0021] Step 3.2, RGB Device Wake-up. Based on Step 3.1, if the results all indicate the existence of the set target intrusion, at this time, the output result of the model is set to 1, and this result is converted into an interrupt output high-frequency start pulse; this pulse can wake up the RGB device to make it turn on for the final verification work.

[0022] Preferably, in Step 3, it is assumed that the model makes a model classification judgment once within T time and outputs whether there is an observed target intrusion. The value range of T is 0.1 s - 0.5 s.

[0023] Step 3, Step 4 further includes:

[0024] Step 4.1, RGB Device Power-on and YOLO Target Detection. Based on Steps 1 to 3, it has been探明 that there is a suspicious intrusion target. At this time, the RGB device starts to capture images, and the current images are sent to the YOLO detection model to determine whether there are pedestrian, vehicle, and drone targets.

[0025] Step 4.2, Evidence Preservation. If the results output by the YOLO detection model include one or more of pedestrians, vehicles, and drones, it indicates that there is indeed a target intrusion in this area. The images are saved in the SD card built into the RGB device and remotely dialed for alarm via 4G. After completing the above work, the RGB device shuts down; if the detection result is other categories, it indicates a false alarm, and at this time, no work is carried out, and the RGB device directly shuts down.

[0026] The present invention also provides a low-power intrusion detection and identification system based on brain-like intelligence, which is applied to the low-power intrusion detection and identification technology based on brain-like intelligence. It consists of a sensing part, a computing part, a storage part, and a 4G alarm part:

[0027] (1) The sensing part includes a DVS camera and an RGB visible light camera. The DVS camera is mainly applied to Steps 1 and 2 for dynamic target capture in the detection area; the RGB visible light camera is mainly applied to Step 4 for intrusion secondary verification image capture.

[0028] (2) The computing part includes an AI computing chip and a brain-like SOC computing chip. The brain-like SOC computing chip is mainly used in step 2 to perform preliminary intrusion detection on the DVS camera field of view. If the output result is a positive sample, the chip will interrupt and output a high-frequency pulse for RGB part startup and wake-up. The AI ​​computing chip is mainly used in step 4 to perform YOLO detection on the current area to confirm whether there is a target intrusion and the intrusion category.

[0029] (3) The storage part and 4G alarm part are mainly used in step 4. When it is determined that there is a target intrusion, the platform saves the evidence of the current intrusion to the SD memory card and remotely alarms.

[0030] The present invention provides a low-power intrusion detection and identification method based on brain-like intelligence. The average power consumption is less than 2mW when there is no foreign object intrusion, and the average power consumption is less than 40mW when intrusion occurs. It can also realize remote 4G wireless alarm. Equipped with a battery, it can realize long-term stable intrusion detection tasks. Thus, it can realize long-term (>20 days) intrusion detection of pedestrians, vehicles, and drones invading a set area with extremely low power consumption, and retain evidence of the intrusion. This technology is applicable to the field of urban digital security, can effectively improve the development of the urban security field, and provide lasting convenience for urban security workers.

[0031] The present invention also provides a computer-readable storage medium, and the computer program, when executed by a processor, performs the steps of low-power intrusion detection and identification technology based on brain-like intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flowchart of the low-power intrusion detection and identification technology based on brain-like intelligence of the present invention;

[0033] Figure 2 This is the drone, vehicle, and pedestrian DVS camera dataset of the present invention;

[0034] Figure 3 It is a schematic diagram of converting an event set within a 3t time range into a frame according to the present invention;

[0035] Figure 4 This is a structural diagram of the pulse convolutional neural network Dynap-CNN model of the present invention;

[0036] Figure 5 It is a schematic diagram of the overall structure of the operating platform of the present invention;

[0037] Figure 6 It is the DVS field of view when there is no intrusion and intrusion on the DVS side, the DVS equipment energy consumption and the preliminary detection alarm diagram;

[0038] Figure 7 This is the preliminary detection result of drone intrusion;

[0039] Figure 8 It is the preliminary detection result of vehicle intrusion;

[0040] Fig. 9 is the system power consumption when there is no pedestrian intrusion;

[0041] Fig.10 is the system power consumption when there is pedestrian intrusion. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0044] In view of the defects and shortcomings of the above-mentioned existing technical solutions, the present invention proposes a low-power intrusion detection and identification solution based on brain-like intelligence, which can detect pedestrians, vehicles, and drones that invade a set area with low power consumption and retain evidence of the intrusion. This solution is applicable to the field of urban digital security, can effectively improve the development of the urban security field, and provide lasting convenience for urban security department staff.

[0045] Example 1

[0046] The technical solution process of the present invention is as follows Figure 1 As shown, the specific steps include:

[0047] Step 1: Dataset processing and data processing enhancement.

[0048] This solution uses dynamic event data captured by DVS cameras as the source of intrusion features. There is no corresponding data set publicly available on the Internet, so pedestrian, vehicle, and drone data must be collected and classified by ourselves. Considering the occlusion and sunlight interference in urban security scenes, we use methods such as image inversion, brightness adjustment, random cropping, edge filling, saturation modification, and MIXUP to enhance the data and improve the robustness of the model.

[0049] Step 2: Build a CNN classification model based on SNN, train it, and deploy it.

[0050] The CNN binary classification model of the intrusion detection SNN architecture is trained on the processed and enhanced dataset to detect whether the target intrusion occurs in the set area.

[0051] Step 3: Intrusion detection and identification and RGB device wake-up.

[0052] Based on step 2, the target within the field of view of the DVS camera is identified. If the recognition result of the classification model is a positive sample, a high-frequency pulse is emitted to wake up the RGB device for secondary detection verification. If the recognition result of the classification model is a negative sample, it means that there is no target intrusion and detection continues.

[0053] Step 4: RGB device identification preserves intrusion evidence.

[0054] Based on steps 1 to 3, the RGB device captures images and sends the captured images to the YOLO series detection model. If the detection result is determined to be a pedestrian, vehicle, or drone, the result is saved and a dial-up intrusion alarm is issued through the 4G module. The RGB device is shut down and waits for the next wake-up. If the detection result is other categories, it is a false alarm, and the RGB device is shut down and waits for the next wake-up.

[0055] Wherein, the step 1 may specifically include the following steps:

[0056] Step 1.1: Data collection, including drone, vehicle, pedestrian and other category datasets.

[0057] For the above dataset, the three categories of drones, vehicles, and pedestrians are taken as positive samples, such as Figure 2 As shown; other categories of data sets include background, various other situations, etc. as negative samples, and the data sets are split into training set, validation set and test set in a ratio of 8:1:1.

[0058] Step 1.2: Data preprocessing.

[0059] The event information output by the DVS camera needs to be converted into image frame information (128*128*2) that can be processed by the brain-like SOC chip model. Event information is generally expressed in the form of a quaternion (x i ,y i ,t i ,p i ), where x i is the horizontal coordinate position of the ith event, y i is the ordinate position of the event, t i is the timestamp of the occurrence of the i-th event, p iIndicates the polarity of the event, generally taking the value of ±1, when the event is enhanced, it takes the value of +1, when the event is weakened, it takes the value of -1. There are generally two methods for converting event sets into frames: segmentation based on a fixed time window and segmentation based on a fixed number of events. Considering the changes in the posture, position, speed, and size of the observed target relative to the DVS camera lens, it is more reasonable to use a segmentation method based on a fixed time window, that is, to divide the event stream according to a fixed time interval, such as Figure 3 As shown. The advantage of this method is that it is simple to operate, but the disadvantage is that it is sensitive to the length of the time window and the speed of movement. In order to reduce the impact of these factors, it is necessary to select a suitable window interval. After experimental verification, the model achieves better results when the window interval is 100ms.

[0060] Step 1.3, dataset enhancement.

[0061] Including image inversion, brightness adjustment, random cropping, edge filling, saturation modification, mixup (where the formula is Img = Img1*0.9+Img2*0.1, note that if Img1 and Img2 are of the same category, the label remains unchanged; if they are of opposite categories, the label needs to be corrected to a positive sample after data enhancement). For the sorted training set, a certain probability is set to randomly select a part of the data, and the above image enhancement methods are added. No data enhancement is performed on the test set.

[0062] Wherein, the step 2 may specifically include the following steps:

[0063] The preprocessed frame (128×128×2) needs to be further processed by the pulse convolutional neural network Dynap-CNN to obtain the category output of the current frame, such as Figure 4 As shown. The data frame will pass through the Conv module, IAFSqueeze module, Pool module, and FC fully connected module in turn to obtain the predicted category, and the predicted category will be cross-entropy calculated with the true category. The weight of the above model is optimized through back propagation, and it is continuously iterated until the model reaches the set accuracy.

[0064] The Dynap-CNN model structure mainly includes a ConV module, an IAFSqueeze module, a Pool module and a FC (fully connected layer) module.

[0065] Step 2.1, ConV module: extract the features that the convolution kernel wants to extract through the convolution kernel operation in the convolution layer. Dynap-CNN includes 3 Conv modules. The number of convolution kernels in the first Conv module is 8, the size of the convolution kernel is 5×5, and the step length is 2; the number of convolution kernels in the second Conv module is 16, the size of the convolution kernel is 5×5, and the step length is 2; the number of convolution kernels in the third Conv module is 8, the size of the convolution kernel is 5×5, and the step length is 1.

[0066] Step 2.2, IAFSqueeze module: A spike neuron model whose main functions are integration and pulse emission. This module integrates by receiving input pulses. Whenever an input pulse arrives, the module integrates it into the internal state variable. This internal state variable can be regarded as the neuron membrane potential, which gradually increases over time. This integration behavior allows neurons to respond to continuous changes in input, rather than just instantaneous input responses. When the internal state variable (neuron membrane potential) exceeds a certain threshold, the module emits an output pulse. This output pulse can be passed to other neurons as their input. The act of emitting pulses allows neurons to transmit and interact with each other through pulses, similar to the pulse transmission in biological neural networks. Before passing through this module, it is necessary to ensure that the 0th dimension of the input data can be expressed as batch size × time length to facilitate subsequent pulse integration and emission. The Dynap-CNN model includes 4 IAFSqueeze modules.

[0067] Step 2.3, Pool module: can reduce the size of spatial information, improve computational efficiency, reduce parameters, reduce the risk of model overfitting, and obtain spatial transformation invariance. Dynap-CNN includes 3 Pool modules.

[0068] Step 2.4, FC (fully connected layer) module: plays the role of "classifier" in the entire convolutional neural network, mapping the feature vector after the output feature map of the last Conv module is expanded into a 2D feature vector.

[0069] Step 2.5 Loss calculation and back propagation: The loss function generally used for binary classification problems is the cross entropy loss function.

[0070]

[0071] Where M is the number of categories, and in this task M is 2. iC is a sign function (0 or 1), if the true category of sample i is equal to c, it takes 1, otherwise it takes 0. iC is the predicted probability that the observed sample i belongs to category c.

[0072] Among them, step 3 can specifically include the following steps:

[0073] Step 3.1, model post-processing and detection recognition.

[0074] The classification output of the Dynap-CNN model is a probability value P between 0 and 1. By setting a threshold S, if P > S, it indicates that there is a potential target intrusion; if P < S, it indicates that there is no intrusion detection. In addition, to improve the accuracy of model classification, a post-processing method of consecutive hits is used. Suppose the model makes a model classification judgment once within time T and outputs whether there is an observed target intrusion. Different from the previous method of alarming as long as there is an output of an observed target intrusion once, now in five consecutive judgments, if there are 4 or more outputs of an observed intrusion, an alarm will be given. This means that the time for one alarm changes from T to 5T, sacrificing some real-time performance but greatly improving the accuracy of the model. Usually, the value of T is 0.1s, and for most scenarios, a judgment time of 0.5s is acceptable.

[0075] Step 3.2, RGB device wake-up.

[0076] Based on step 3.1, if the results all indicate the existence of a set target intrusion, at this time, the output result of the model is set to 1, and this result is converted into an interrupt output high-frequency start pulse; this pulse can wake up the RGB device to turn it on for the final verification work.

[0077] Among them, step 4 can specifically include the following steps:

[0078] Step 4.1, RGB device startup and YOLO target detection: Based on steps 1 to 3, it has been determined that there is a suspicious intrusion target. At this time, the RGB device starts to capture images. The current image is sent to the YOLO detection model to determine whether there are pedestrian, vehicle, and drone targets.

[0079] Step 4.2, evidence preservation: If the results output by the YOLO detection model include one or more of pedestrians, vehicles, and drones, it indicates that there is indeed a target intrusion in this area. The image is saved in the SD card built into the RGB device and an alarm is sent via remote 4G dialing. After completing the above work, the RGB device shuts down; if the detection result is other categories, it indicates a false alarm, and at this time, no work is done and the RGB device directly shuts down.

[0080] Based on the above solution, the operating platform of the present invention is as Figure 5 shown and mainly consists of four parts:

[0081] (1) The perception part includes DVS camera and RGB visible light camera. The DVS camera is mainly used in steps 1 and 2 to capture dynamic targets in the detection area; the RGB visible light camera is mainly used in step 4 to capture intrusion secondary verification images;

[0082] (2) The computing part includes AI computing chips and brain-like SOC computing chips. The brain-like SOC computing chip is mainly used in step 2 to perform preliminary intrusion detection on the DVS camera field of view. If the output result is a positive sample, the chip will interrupt and output high-frequency pulses for RGB part power-on wake-up. The AI ​​computing chip is mainly used in step 4 to perform YOLO detection on the current area to confirm whether there is a target intrusion and the intrusion category.

[0083] (3) The storage part and 4G alarm part are mainly used in step 4. When it is determined that there is a target intrusion, the platform saves the evidence of the current intrusion to the SD memory card and remotely alarms.

[0084] The reconstruction technology used in the present invention brings the following advantages:

[0085] The dual-mode integrated intrusion detection solution designed by the present invention has an average power consumption of less than 2mW when there is no foreign object intrusion, and an average power consumption of less than 40mW when there is intrusion, and can realize remote 4G wireless alarm. It is equipped with a battery to achieve long-term and stable intrusion detection tasks.

[0086] Example 2

[0087] In this embodiment, the Dynap-CNN classification model based on SNN can be used to perform low-power discrimination on targets in the detection area. The relevant parameters of the model are shown in Table 1. The classification results calculated by the trained model for a single category are shown in Table 2.

[0088] Table 1 Model parameters

[0089] Serial number category Value 1 Data preprocessing time setting (ms) 100 2 Model Layers 10 3 Frame size (FRAME) 128*128*2 4 Initial learning rate (lr) 0.001 5 Batch size 64 6 Training round (epoch) 50 7 Random loss rate (sigout) 0.1 8 Is noise reduction enabled? True 9 Learning rate adjustment strategy (lr adjust type) MultiStepLR

[0090] The initial learning rate is set to 0.001, and the learning rate adjustment strategy is MultiStepLR. The learning rate can be adjusted between the 5th and 30th rounds of model training, and each adjustment is 98% higher or lower than the previous learning rate. The specific adjustment formula is as follows:

[0091]

[0092] When the noise reduction is set to True, the signal can be filtered and noise reduced when it is input.

[0093] The random dropout rate is set to 0.1, which means that 10% of the data before entering the model can be randomly lost, that is, the pixel value of 12*12 in the image is set to 0. This parameter is derived from the idea of ​​Dropout in deep learning. After many experiments, it is found that setting the random dropout rate to 0.1 has the best effect.

[0094] Table 2 Detection results based on DVS brain-like detection model

[0095] category Recognition rate Accuracy Recall F1 pedestrian 92.60％ 95.60％ 92.60％ 94.07％ vehicle 94.50％ 95.50％ 91.50％ 93.45％ Drones 94.50％ 96.50％ 93.50％ 94.97％

[0096] In order to verify the accuracy and low power consumption of the intrusion detection alarm of the system, several experiments on target detection and intrusion alarm were carried out. Figure 6-10 As shown, Figure 6 The left side shows the total power consumption of the DVS field of view and detection system when there is no intrusion (see Fig. 9 Detailed power consumption diagram), Figure 6 The right side shows the total power consumption of the DVS field of view and detection system when the target invades (see Fig.10 Detailed power consumption diagram). Figure 7 The detection system's initial identification results when a drone intrudes. Figure 8 This is the initial recognition result of the detection system when a vehicle intrudes. It can be found that when pedestrians, vehicles, and drones enter the field of view of the DVS device, the SOC brain-like chip runs the Dynap-CNN algorithm to enter the initial detection, verifying the initial detection accuracy and low power consumption characteristics of the system.

[0097] In order to verify the low power consumption feature of the system, the system power consumption is recorded when there is no intrusion and when there is target intrusion. Fig. 9 The system power consumption when there is no intrusion is recorded. Fig.10 The system power consumption when there is a pedestrian intrusion is recorded. The total power consumption of the system is the sum of the power consumption of the five parts "io", "ram", "logic", "vddd" and "vdda", and the unit of the vertical axis is mW. When there is no pedestrian intrusion, the total power consumption of the system is less than 1.5mW, and when there is a pedestrian intrusion, the total power consumption of the system is less than 20mW, which verifies the low power consumption characteristics of the system.

[0098] In another embodiment, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that when the computer program is executed by a processor, a low-power intrusion detection and identification technical step based on brain-like intelligence is implemented.

[0099] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory.

[0100] Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0101] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0102] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. A low-power intrusion detection and identification technology based on brain-like intelligence, characterized in that: include: Step 1: Data set processing and data processing enhancement; Step 2: Build and deploy a CNN classification model based on SNN, and train a CNN binary classification model of the intrusion detection SNN architecture on the processed and enhanced data set to detect whether a target intrusion occurs in the set area; Step 3: Intrusion detection and recognition and RGB device wake-up. Based on step 2, the target in the field of view of the DVS camera is identified. If the classification model recognizes a positive sample, a high-frequency pulse is emitted to wake up the RGB device for secondary detection verification. If the classification model recognizes a negative sample, it means that there is no target intrusion and detection continues. Step 4: The RGB device identifies and retains the intrusion evidence. Based on steps 1 to 3, the RGB device captures the image and sends the captured image to the YOLO detection model. If the detection result is determined to be a positive sample of the target intrusion, the result is saved and a dial-up intrusion alarm is issued through the communication module. The RGB device is shut down and waits for the next wake-up. If the detection result is a negative sample, it is a false alarm, and the RGB device is shut down and waits for the next wake-up.

2. The technology of claim 1, wherein: The step 1 further comprises: Step 1.1, data collection, including drone, vehicle, pedestrian and other category data sets, drone, vehicle, pedestrian three categories as positive samples, other category data sets including background, various other situations as negative samples, and the data set is split into training set, validation set and test set in the ratio of 8:1:1; Step 1.2: Data preprocessing: converting the event information output by the DVS camera into image frame information that can be processed by the brain-like SOC chip model; Step 1.3, data set enhancement, using one or more methods such as image inversion, brightness adjustment, random cropping, edge filling, saturation modification, mixup, etc. to enhance the data set; for the sorted training set, set a certain probability to randomly select a part of the data, add the above image enhancement method, and do not perform any data enhancement on the test set.

3. The technology of claim 1, wherein: The step 2 further includes: The frames after data preprocessing are further processed by the pulse convolutional neural network Dynap-CNN to obtain the category output of the current frame. The data frame will pass through the Conv module, IAFSqueeze module, Pool module, and FC (fully connected layer) in sequence to obtain the predicted category. The cross entropy loss is calculated between the predicted category and the true category, and the weights of the above model are optimized through back propagation, and it is continuously iterated until the model reaches the set accuracy.

4. The technique of claim 3, wherein: The Dynap-CNN includes 3 Conv modules. The number of convolution kernels in the first Conv module is 8, the size of the convolution kernel is 5×5, and the step size is 2; the number of convolution kernels in the second Conv module is 16, the size of the convolution kernel is 5×5, and the step size is 2; the number of convolution kernels in the third Conv module is 8, the size of the convolution kernel is 5×5, and the step size is 1.

5. The technique of claim 1, wherein: The step 3 further includes: Step 3.1, Model Post-Processing and Detection Recognition. The classification output of the Dynap-CNN model is a probability value P ranging from 0 to 1. By setting a threshold S, if P > S, it indicates that there is a potential target intrusion; if P < S, it indicates that there is no intrusion detection. Step 3.2, RGB Device Wake-up. Based on Step 3.1, if the results all indicate the existence of a set target intrusion, at this time, the output result of the model is set to 1, and this result is converted into an interrupt output high-frequency start pulse; this pulse can wake up the RGB device to turn it on for the final verification work.

6. The technique of claim 5, wherein: It is assumed that the model makes a model classification judgment once within time T, and outputs whether there is an observed target intrusion. The value range of T is 0.1s - 0.5s.

7. The technique of claim 1, wherein: Said Step 4 further includes: Step 4.1, RGB Device Boot-up and YOLO Target Detection. Based on Steps 1 to 3, it has been determined that there is a suspicious intrusion target. At this time, the RGB device starts to capture images, and the current images are sent to the YOLO detection model to determine whether there are pedestrian, vehicle, and drone targets. Step 4.2, Evidence Preservation. If the results output by the YOLO detection model include one or more of pedestrians, vehicles, and drones, it indicates that there is indeed a target intrusion in this area. The images are saved in the SD card built into the RGB device and remotely dialed for alarm via 4G. After completing the above work, the RGB device shuts down; if the detection result is other categories, it indicates a false alarm, and at this time, no work is carried out, and the RGB device directly shuts down.

8. A low-power intrusion detection and identification system based on brain-like intelligence, applied to the technology described in claims 1-5, characterized in that: It consists of a sensing part, a computing part, a storage part, and a 4G alarm part: (1) The sensing part includes a DVS camera and an RGB visible light camera. Among them, the DVS camera is mainly used in Steps 1 and 2 to capture dynamic targets in the detection area; the RGB visible light camera is mainly used in Step 4 to capture images for secondary verification of intrusion. (2) The computing part includes an AI computing chip and a brain-like SOC computing chip. Among them, the brain-like SOC computing chip is mainly used in Step 2 to conduct a preliminary intrusion detection on the field of view of the DVS camera. If the output result is a positive sample, this chip will interrupt and output a high-frequency pulse for waking up the RGB part to turn it on; the AI computing chip is mainly used in Step 4 to conduct a YOLO detection on the current area to confirm whether there is a target intrusion and the intrusion category. (3) The storage part and the 4G alarm part are mainly used in Step 4. When it is determined that there is a target intrusion, the platform saves the evidence of the current intrusion to the SD memory card and remotely alarms.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the low-power intrusion detection and recognition technology based on brain-like intelligence described in any one of Claims 1 to 7.

Images