A multi-scale anomalous signal detection method for infrared signals

Abstract

A multi-scale anomaly signal detection method for infrared signals, relating to the field of infrared signal detection, includes: acquiring infrared signal data; initializing infrared signal sequences; constructing multi-scale infrared signal data; background modeling of the multi-scale infrared signal data; and multi-scale anomaly signal detection. This invention eliminates the need for manually setting thresholds, reducing false triggering caused by threshold incompatibility; it enables adaptive targeting at different scales and motion speeds through multi-scale data, and Gaussian filtering reduces signal noise; it estimates background data distribution through background modeling, allowing for the simulation of more complex scenarios using multiple modes; it is robust to changes in lighting, target swaying, and equipment shaking, improving noise resistance; and it is applicable to any moving target without requiring an additional target sample library.

Description

Technical Field

[0001] This invention relates to the field of infrared signal detection technology, and more specifically to a method for detecting multi-scale abnormal signals in infrared signals. Background Technology

[0002] Infrared sensors can collect energy reflected from sunlight or electromagnetic waves radiated by ground targets. They sense this electromagnetic energy and output a thermoelectric signal, which, after amplification and clutter removal, can be used for target detection. While infrared sensors cannot directly provide distance information, their thermal radiation sensitivity and excellent concealment make them widely used in electronic security and human detection systems.

[0003] Existing infrared target detection methods generally rely on a set threshold, assuming the infrared signal in the environment remains relatively constant. During detection, adjacent signals are differentially analyzed, and an alarm is triggered when the intensity of the differential signal exceeds the threshold. However, these methods typically only consider indoor environments. In real-world environments, such as complex outdoor scenarios, external interference signals, such as changes in lighting or swaying tree branches, can cause false triggers, impacting user experience. Therefore, existing infrared target detection methods are not applicable to all scenarios. Consequently, developing a robust infrared signal detection method has become a pressing technical problem for those skilled in the art.

[0004] Chinese patent CN114355465A discloses a "Human Body Intelligent Sensing Device and its Infrared Signal Interference Filtering Method." This human body intelligent sensing device mainly includes an infrared signal receiver and a signal processor, with the signal processor connected to the receiver. In this infrared signal interference filtering method, background noise is first filtered out from the interference signal. The infrared signal after background noise filtering is compared with a set threshold. If it exceeds the set threshold, it is considered that someone is approaching the human body intelligent sensing device, and the device is triggered. This method has the following problems: the background modeling method is simple, generally collecting infrared signals without a target as background data, and only performing simple filtering on the collected infrared signals, failing to achieve effective background modeling. Furthermore, it is only effective for simple scenes such as indoors, and in practical applications, it often causes false triggering in complex scenes such as outdoors.

[0005] Chinese patent CN110827499A discloses a "Method and Electronic Device for Detecting Moving Objects." This method determines the comparison feature parameters of an infrared signal based on a preset data processing algorithm; compares these comparison feature parameters with each feature parameter in a state information database; and determines the state information of the moving object based on the successfully matched state information. The method has limited target detection capabilities, primarily because it requires pre-establishing a state information database for specific targets such as flames, pedestrians, and vehicles. By pre-collecting this database and storing the correspondence between various pre-obtained state information and feature parameters, false triggering is possible when a new scene is similar to the database. Furthermore, the state information database cannot be adaptively updated, and ensuring its comprehensive coverage across multiple industries is challenging. Ensuring the completeness of the state information database is a necessary prerequisite; therefore, this method is only applicable to specified targets. Summary of the Invention

[0006] To address the problems of existing methods, such as simple background modeling leading to frequent false triggers, inability to adapt to arbitrary scenarios, and limited target types, this invention provides a multi-scale abnormal signal detection method for infrared signals.

[0007] The technical solution adopted by this invention to solve the technical problem is as follows:

[0008] The present invention provides a method for detecting multi-scale anomaly signals in infrared signals, comprising the following steps:

[0009] Step 1: Acquire infrared signal data;

[0010] Step 2: Initialize the infrared signal sequence;

[0011] Step 3: Construct multi-scale infrared signal data;

[0012] Step 4: Background modeling of multi-scale infrared signal data;

[0013] Step 5: Multi-scale anomaly signal detection.

[0014] Furthermore, in step one, multiple infrared signal data are acquired from the infrared point source detector.

[0015] Furthermore, the specific operation process of step two is as follows:

[0016] The length of the valid data in the acquired infrared signal data is 1000, and the corresponding set of infrared signal sequences is denoted as X. bg =(x1,x2,x3,...,x 999 ,x 1000 ), x1, x2, x3, ..., x 999 ,x 1000Representing the set of infrared signal sequences X bg The infrared signal sequence from time 1 to time 1000 is used to initialize the infrared signal sequence and determine whether the initialization of the infrared signal sequence is complete. If yes, proceed to step 3; otherwise, return to step 1 to reacquire the infrared signal data and reinitialize the infrared signal sequence until the initialization of the infrared signal sequence is complete.

[0017] Furthermore, the specific operation process of step three is as follows:

[0018] S3.1 Set the infrared signal sequence X bg The infrared signal sequence information in the set X is used as historical sequence information for the infrared signal sequence set. bg The infrared signal sequences in the image are subjected to one-dimensional Gaussian smoothing, and then convolved with a convolution kernel of size 5. X bgp (c) represents the set of infrared signal sequences X after one-dimensional Gaussian smoothing. bgp The c-th element, X bg (c) represents the set of infrared signal sequences X bg The c-th element in the convolution kernel, where c ∈ {1, 2, 3, ..., 1000}, has the following kernel:

[0019] S3.2 The set of infrared signal sequences X after one-dimensional Gaussian smoothing bgp The infrared signal sequences in the image are downsampled, and even-numbered columns are removed to obtain the dimension-reduced infrared signal sequences. r represents the downsampling number, r∈{1,2,3};

[0020] S3.3 Dimensionally reduced infrared signal sequence By repeating steps S3.1 to S3.2 three times, multi-scale infrared signal data can be obtained. The set of multi-scale infrared signal data is denoted as... X bg This represents the original set of infrared signal sequences, with a length of 1000; These represent the infrared signal sequences after the first, second, and third dimensionality reduction, respectively, with lengths of 500, 250, and 125.

[0021] Furthermore, the specific operation process for step four is as follows:

[0022] S4.1 For the multi-scale data of a certain infrared signal obtained above H(j) is the j-th element in the multi-scale infrared signal data set H, where j∈{1,2,3,4} and t∈{1,2,...,N}. The probability density function of the mixture Gaussian distribution The calculation formula is as follows:

[0023]

[0024]

[0025]

[0026] In the formula, k is the total number of Gaussian distribution models; i∈{1,2,...,k}; w i,t The weights of the i-th Gaussian distribution model at time t; Let be the i-th Gaussian distribution model at time t. Let μ be the signal quantity at time t at the current scale. i,t Let τ be the mean of the i-th Gaussian distribution model at time t. i,t Let δ be the covariance matrix of the i-th Gaussian distribution model at time t, and T be the transpose operation; i,t Let be the variance of the i-th Gaussian distribution model at time t, and I be the identity matrix;

[0027] S4.2 Initialize the Gaussian distribution model with a total number of k = 3, a learning rate α = 0.05, and initialize 3 Gaussian distribution means and 3 variances;

[0028] S4.3 Acquire all infrared signal data for the first scale one by one, and check whether the current infrared signal data matches the current three Gaussian distribution models. That is, the mean deviation of the matched Gaussian distribution models should be within 2.5δ. k,t-1 within, that is Let μ be the signal quantity at time t at the current scale. k,t-1 Let δ be the mean of the k-th Gaussian distribution model at time t-1. k,t-1 Let $\mathbf{k}$ be the variance of the $k$-th Gaussian distribution model at time $t-1$.

[0029] S4.4 If the matched Gaussian distribution model meets the background requirements, then define M. k,t =1, otherwise M k,t =0;

[0030] The weights of each Gaussian distribution model in S4.5 are updated according to the following formula:

[0031] w k,t = (1-α)*w k,t-1 +α*M k,t

[0032] In the formula, α = 0.05; w k,t w represents the weight of the k-th Gaussian distribution model at time t. k,t-1 The weights of the k-th Gaussian distribution model at time t-1;

[0033] S4.6 The mean and standard deviation of the unmatched Gaussian distribution model remain unchanged, while the parameters of the matched Gaussian distribution model are updated according to the following formula:

[0034]

[0035]

[0036]

[0037] In the formula, ρ is the weight value during iteration; α = 0.05; The model is the k-th Gaussian distribution at time t-1; μ k,t μ is the mean of the k-th Gaussian distribution model at time t; k,t-1 σ is the mean of the k-th Gaussian distribution model at time t-1; k,t σ is the standard deviation of the k-th Gaussian distribution model at time t; k,t-1 Let $\frac{k}{t-1}$ be the standard deviation of the $k$-th Gaussian distribution model at time $t-1$.

[0038] S4.7 If the first infrared signal data at the current scale does not have a matching Gaussian distribution model, the Gaussian distribution model with the smallest weight is replaced. That is, the mean of the Gaussian distribution model is the current pixel value, the standard deviation of the Gaussian distribution model is the initial larger value, and the weight of the Gaussian distribution model is the smaller value.

[0039] S4.8 Gaussian distribution models based on w k,t / σ k,t Arrange them in descending order, that is, Gaussian distribution models with large weights and small standard deviations are arranged first;

[0040] S4.9 Selects the first B Gaussian distribution models as the background, where B satisfies the following equation:

[0041]

[0042] S4.10 Repeat the background modeling process from S4.1 to S4.9 above for the four scale data extracted in step three (S3.3) to complete the background modeling of infrared signal multi-scale data.

[0043] Furthermore, the specific operation process of step five is as follows:

[0044] S5.1 Acquisition of data to be detected;

[0045] Obtain the infrared signal x at the current moment from the infrared point source detector.cur ;

[0046] S5.2 View the current infrared signal x cur Whether it matches B Gaussian distribution models at four scales, i.e., the mean deviation of the matching Gaussian distribution model should be within 2.5δ. k,1000 Inside, i.e., |x cur -μ k,1000 |≤2.5δ k,1000 , where μ k,1000 δ is the mean of the k-th Gaussian distribution model at the last time step in the current scale. k,1000 Let x be the standard deviation of the k-th Gaussian distribution model at the last moment in the current scale, where k∈{1,2,...,B}; if the infrared signal x at the current moment... cur If the signal does not match any of the B Gaussian distribution models at the four scales, then the current signal is determined to contain an abnormal target; otherwise, no abnormal target is found.

[0047] S5.3 Infrared signal sequence update;

[0048] If no abnormal signal is detected in step S5.2, the infrared signal sequence is updated by deleting the infrared signal sequence set X. bg The first infrared signal sequence in the set, and insert the current infrared signal sequence into the infrared signal sequence set X. bg The end of.

[0049] The beneficial effects of this invention are:

[0050] This invention extracts multi-scale information from infrared signals and uses a background modeling algorithm to perform background noise estimation and target detection. Compared with existing technologies, this invention's method for multi-scale anomaly signal detection in infrared signals brings improved effectiveness and convenience to infrared signal-based target detection in the following aspects:

[0051] 1. No need to manually set thresholds, reducing false triggering caused by threshold mismatch.

[0052] 2. Multi-scale data can be used to adapt the target to different scales and motion speeds, and Gaussian filtering can also reduce signal noise.

[0053] 3. Strong noise resistance: Background data distribution is estimated through background modeling, and multiple modes can simulate more complex scenes. It is also robust to changes in lighting, swaying of interfering targets, and shaking of equipment.

[0054] 4. Unbiasedness: The background modeling algorithm only maximizes the probability and does not make the mean biased towards 0 or the cluster size biased towards special structures that may or may not be applicable.

[0055] 5. Fast detection speed: The multi-scale anomaly signal detection method for infrared signals proposed in this invention is the fastest method among hybrid model learning algorithms.

[0056] 6. No limit on target type: This invention is applicable to any moving target and does not require the establishment of an additional target sample library. Attached Figure Description

[0057] Figure 1 This is a flowchart of a multi-scale abnormal signal detection method for infrared signals according to the present invention. Detailed Implementation

[0058] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0059] The present invention provides a multi-scale anomaly signal detection method for infrared signals, which mainly includes the following steps:

[0060] Step 1: Acquire infrared signal data;

[0061] Step 2: Initialize the infrared signal sequence;

[0062] Step 3: Construct multi-scale infrared signal data;

[0063] Step 4: Background modeling of multi-scale infrared signal data;

[0064] Step 5: Multi-scale anomaly signal detection.

[0065] like Figure 1 As shown, the present invention provides a method for detecting multi-scale anomaly signals in infrared signals, which specifically includes the following steps:

[0066] Step 1: Acquire infrared signal data;

[0067] Multiple infrared signal data are acquired from an infrared point source detector.

[0068] Step 2: Initialize the infrared signal sequence;

[0069] The length of the valid data in the acquired infrared signal data is 1000, and the corresponding set of infrared signal sequences is denoted as X. bg =(x1,x2,x3,...,x 999 ,x 1000 ), x1, x2, x3, ..., x 999 ,x1000 Representing the set of infrared signal sequences X bg The infrared signal sequence from time 1 to time 1000 is used to initialize the infrared signal sequence and determine whether the initialization of the infrared signal sequence is complete. If yes, proceed to step 3; otherwise, return to step 1 to reacquire the infrared signal data and reinitialize the infrared signal sequence until the initialization of the infrared signal sequence is complete.

[0070] Step 3: Construct multi-scale infrared signal data;

[0071] Multi-scale data refers to infrared signal data at different temporal resolutions, used to adapt to targets moving at different speeds. The specific operational process of this step is as follows:

[0072] S3.1 Set the infrared signal sequence X bg The infrared signal sequence information in the set X is used as historical sequence information for the infrared signal sequence set. bg The infrared signal sequences in the image are subjected to one-dimensional Gaussian smoothing. Specifically, a convolution kernel of size 5 can be used to perform a convolution operation on the infrared signal sequences. Among them, X bgp (c) represents the set of infrared signal sequences X after one-dimensional Gaussian smoothing. bgp The c-th element (referring to the infrared signal sequence set X) bgp (the c-th infrared signal sequence in the sequence), X bg (c) represents the set of infrared signal sequences X bg The c-th element (referring to the infrared signal sequence set X) bg The c-th infrared signal sequence in the sequence, c∈{1,2,3,……999,1000}, has a convolution kernel of...

[0073] S3.2 The set of infrared signal sequences X after one-dimensional Gaussian smoothing bgp The infrared signal sequences in the image are downsampled, and even-numbered columns are removed to obtain the dimension-reduced infrared signal sequences. Where r represents the number of downsampling operations, r∈{1,2,3};

[0074] S3.3 Dimensionally reduced infrared signal sequence By repeating steps S3.1 to S3.2 three times, multi-scale infrared signal data can be obtained. The set of multi-scale infrared signal data is denoted as... Among them, X bg This represents the original set of infrared signal sequences, with a length of 1000; These represent the infrared signal sequences after the first, second, and third dimensionality reduction, respectively, with lengths of 500, 250, and 125.

[0075] Step 4: Background modeling of multi-scale infrared signal data;

[0076] Background modeling refers to characterizing the distribution of infrared signals in the absence of a target by using statistical information such as the probability density of a large number of sample values ​​over a long period of time. Background modeling is required for each data point in multi-scale infrared signal data. Through background modeling, the change of the infrared signal sequence over time can be viewed as a random process that continuously generates signal values. This random process conforms to a multi-peak Gaussian distribution model. The infrared signal sequence at any given time can be modeled using the superposition of multiple Gaussian distribution models with different weights. Each Gaussian distribution model corresponds to a possible state that could generate an infrared signal sequence, and the weights and distribution parameters of each Gaussian distribution model are updated over time.

[0077] The specific steps for background modeling are as follows:

[0078] S4.1 For the multi-scale data of a certain infrared signal obtained above H(j) is the j-th element in the infrared signal multi-scale data set H (the j-th element refers to the j-th infrared signal multi-scale data in the infrared signal multi-scale data set H), j∈{1,2,3,4}, t∈{1,2,...,N}, then The probability density function of the mixture Gaussian distribution The calculation formula is as follows:

[0079]

[0080]

[0081]

[0082] In the formula, k is the total number of Gaussian distribution models; i∈{1,2,...,k}; w i,t The weights of the i-th Gaussian distribution model at time t; Let be the i-th Gaussian distribution model at time t, where Let μ be the signal quantity at time t at the current scale. i,t Let τ be the mean of the i-th Gaussian distribution model at time t. i,t Let δ be the covariance matrix of the i-th Gaussian distribution model at time t, and T be the transpose operation; i,t Let be the variance of the i-th Gaussian distribution model at time t, and I be the identity matrix;

[0083] S4.2 Initialize the Gaussian distribution model with a total number of k = 3, a learning rate α = 0.05, and initialize 3 Gaussian distribution means and 3 variances;

[0084] S4.3 Acquire all infrared signal data for the first scale one by one, and check whether the current infrared signal data matches the current three Gaussian distribution models. That is, the mean deviation of the matched Gaussian distribution models should be within 2.5δ. k,t-1 within, that is in, Let μ be the signal quantity at time t at the current scale. k,t-1 Let δ be the mean of the k-th Gaussian distribution model at time t-1. k,t-1 Let $\mathbf{k}$ be the variance of the $k$-th Gaussian distribution model at time $t-1$.

[0085] S4.4 If the matched Gaussian distribution model meets the background requirements, then define M. k,t =1, otherwise M k,t =0;

[0086] The weights of each Gaussian distribution model in S4.5 are updated according to the following formula:

[0087] w k,t = (1-α)*w k,t-1 +α*M k,t

[0088] In the formula, α = 0.05; w k,t w represents the weight of the k-th Gaussian distribution model at time t. k,t-1 The weights of the k-th Gaussian distribution model at time t-1;

[0089] S4.6 The mean and standard deviation of the unmatched Gaussian distribution model remain unchanged, while the parameters of the matched Gaussian distribution model are updated according to the following formula:

[0090]

[0091]

[0092]

[0093] In the formula, ρ is the weight value during iteration; α = 0.05; The model is the k-th Gaussian distribution at time t-1; μ k,t μ is the mean of the k-th Gaussian distribution model at time t; k,t-1 σ is the mean of the k-th Gaussian distribution model at time t-1; k,t σ is the standard deviation of the k-th Gaussian distribution model at time t; k,t-1 Let $\frac{k}{t-1}$ be the standard deviation of the $k$-th Gaussian distribution model at time $t-1$.

[0094] S4.7 If the first infrared signal data at the current scale does not have a matching Gaussian distribution model, the Gaussian distribution model with the smallest weight is replaced. That is, the mean of the Gaussian distribution model is the current pixel value, the standard deviation of the Gaussian distribution model is the initial larger value, and the weight of the Gaussian distribution model is the smaller value.

[0095] S4.8 Gaussian distribution models based on w k,t / σ k,t Arrange them in descending order, that is, Gaussian distribution models with large weights and small standard deviations are arranged first;

[0096] S4.9 Selects the first B Gaussian distribution models as the background, where B satisfies the following equation:

[0097]

[0098] S4.10 refers to the four scale data extracted in step three, S3.3. Repeat the background modeling process from S4.1 to S4.9 above to complete the background modeling of infrared signal multi-scale data.

[0099] Step 5: Multi-scale anomaly signal detection;

[0100] S5.1 Acquisition of data to be detected;

[0101] Obtain the infrared signal x at the current moment from the infrared point source detector. cur ;

[0102] S5.2 View the current infrared signal x cur Whether it matches B Gaussian distribution models at four scales, i.e., the mean deviation of the matching Gaussian distribution model should be within 2.5δ. k,1000 Inside, i.e., |x cur -μ k,1000 |≤2.5δ k,1000 , where μ k,1000 δ is the mean of the k-th Gaussian distribution model at the last time step in the current scale. k,1000 Let x be the standard deviation of the k-th Gaussian distribution model at the last moment in the current scale, where k∈{1,2,...,B}; if the infrared signal x at the current moment... cur If the signal does not match any of the B Gaussian distribution models at the four scales, then the current signal is determined to contain an abnormal target; otherwise, no abnormal target is found.

[0103] S5.3 Infrared signal sequence update;

[0104] If no abnormal signal is detected in step S5.2, the infrared signal sequence is updated by deleting the infrared signal sequence set X. bgThe first infrared signal sequence in the set, and insert the current infrared signal sequence into the infrared signal sequence set X. bg The end of.

[0105] Step 6: For subsequent new infrared signals, repeat steps 1 to 5 to achieve multi-scale anomaly signal detection of infrared signals.

[0106] This invention provides a multi-scale anomaly signal detection method for infrared signals. This method models the background signal at multiple scales, learning the background at each scale using statistical information such as the probability density of a large number of sample values ​​over a long period. This allows for modeling of complex dynamic backgrounds and achieving target detection. The method is applicable to complex indoor and outdoor scenes and can effectively model interference such as changes in lighting, swaying tree branches, or equipment vibration.

[0107] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multiscale anomaly signal detection method for infrared signals, characterized in that, The method comprises the following steps: Step one, acquiring infrared signal data; acquiring a plurality of infrared signal data from an infrared point source detector; Step two, initializing an infrared signal sequence; The length of the valid data in the acquired infrared signal data is 1000, and the corresponding set of infrared signal sequences is denoted as X. bg =(x1,x2,x3,...,x 999 ,x 1000 ), x1, x2, x3, ..., x 999 ,x 1000 Representing the set of infrared signal sequences X bg The infrared signal sequence from the 1st to the 1000th time step is used to initialize the infrared signal sequence and determine whether the initialization of the infrared signal sequence is complete. If yes, proceed to step three; otherwise, return to step one to reacquire the infrared signal data and reinitialize the infrared signal sequence until the initialization of the infrared signal sequence is complete. Step three, constructing infrared signal multi-scale data; S3.1 Take the infrared signal sequence information in the infrared signal sequence set X bg as historical sequence information, and do one-dimensional Gaussian smoothing processing on the infrared signal sequences in the infrared signal sequence set X bg respectively, using a convolution kernel with a size of 5 to perform convolution operation on the infrared signal sequences: X bgp (c) represents the cth element in the infrared signal sequence set X bgp after one-dimensional Gaussian smoothing processing, X bg (c) represents the cth element in the infrared signal sequence set X bg , c ∈ {1, 2, 3, …, 1000}, and the convolution kernel is S3.2 down-sampling the infrared signal sequences in the set X bgp to obtain the reduced dimension infrared signal sequences r represents the number of down-sampling, r∈{1,2,3}. S3.3 The infrared signal sequence after dimension reduction Steps S3.1 to S3.2 are repeated three times to obtain infrared signal multiscale data, which is denoted as X bg represents the original infrared signal sequence set, with a length of 1000; respectively represent the first, second and third infrared signal sequences after dimension reduction, with lengths of 500, 250 and 125, respectively. Step four, background modeling of infrared signal multi-scale data; S4.1 For the above obtained certain infrared signal multiscale data H(j) is the jth element in the set of infrared signal multiscale data H, j ∈ {1, 2, 3, 4}, t ∈ {1, 2, …, N}, then The formula for calculating the probability density function of the Gaussian mixture distribution subject to The formula for calculating the probability density function of the Gaussian mixture distribution subject to wherein k is the total number of Gaussian distribution models; i ∈ {1, 2, …, k}; w i,t is the weight of the i-th Gaussian distribution model at time t; is the i-th Gaussian distribution model at time t, is the signal quantity at time t under the current scale, μ i,t is the mean of the i-th Gaussian distribution model at time t, τ i,t is the covariance matrix of the i-th Gaussian distribution model at time t, T is the transpose operation; δ i,t is the variance of the i-th Gaussian distribution model at time t, I is the unit matrix; S4.2 initializing the total number of Gaussian distribution models k as 3, the learning rate a as 0.05, initializing the mean values of the three Gaussian distribution models and the three variances; S4.3 All infrared signal data of the first scale is acquired one by one, and it is checked whether the current infrared signal data matches the current three Gaussian distribution models, i.e. the deviation of the matching Gaussian distribution model mean value should be within 2.5δ k,t-1 , i.e. is the signal amount at time t under the current scale, μ k,t-1 is the mean value of the kth Gaussian distribution model at time t-1, δ k,t-1 is the variance of the kth Gaussian distribution model at time t-1; S4.4 If the matched Gaussian distribution model meets the background requirement, define M k,t = 1, otherwise M k,t = 0; S4.5 the weight of each Gaussian distribution model is updated according to the following formula: w k,t = (1 - a) * w k,t-1 + a * M k,t wherein a = 0.05; w k,t is the kth Gaussian distribution model weight at time t; w k,t-1 is the kth Gaussian distribution model weight at time t-1; S4.6 the mean value and the standard deviation of the unmatched Gaussian distribution model are unchanged, and the parameters of the matched Gaussian distribution model are updated according to the following formula: wherein p is the weight value at iteration; a = 0.05; is the kth Gaussian distribution model at time t-1; μ k,t is the mean of the kth Gaussian distribution model at time t; μ k,t-1 is the mean of the kth Gaussian distribution model at time t-1; σ k,t is the standard deviation of the kth Gaussian distribution model at time t; σ k,t-1 is the standard deviation of the kth Gaussian distribution model at time t-1; S4.7 if the first infrared signal data of the current scale cannot be matched with any Gaussian distribution model, the Gaussian distribution model with the smallest weight is replaced, that is, the mean value of the Gaussian distribution model is the current pixel value, the standard deviation of the Gaussian distribution model is an initial large value, and the weight of the Gaussian distribution model is a small value; S4.8 Each Gaussian distribution model is assigned a weight w k,t / σ k,t The Gaussian distribution models are arranged in descending order, i.e. the Gaussian distribution models with large weights and small standard deviations are arranged first. S4.9 selecting the first B Gaussian distribution models as the background, wherein B satisfies the following formula: S4.10 repeating the background modeling process of S4.1 to S4.9 above for the four scale data extracted in S3.3 in step three, to complete the background modeling of the infrared signal multi-scale data; Step five, multi-scale abnormal signal detection; S5.1 acquiring the data to be detected; Obtaining the infrared signal x at the current time from the infrared point source detector cur ; S5.2 Check the current time infrared signal x cur whether matches the B Gaussian distribution models under 4 scales, that is, the matching Gaussian distribution model mean deviation should be within 2.5δ k,1000 , that is, |x cur -μ k,1000 |≤2.5δ k,1000 , where μ k,1000 is the kth Gaussian distribution model mean of the last time under the current scale, δ k,1000 is the kth Gaussian distribution model standard deviation of the last time under the current scale, k∈{1,2,...,B}; if the current time infrared signal x cur does not match the B Gaussian distribution models under 4 scales, it is judged that there is an abnormal target in the current signal, otherwise there is no abnormal target. S5.3 updating the infrared signal sequence; When no abnormal signal is detected in step S5.2, then the infrared signal sequence is updated in a way that the first infrared signal sequence in the set X of infrared signal sequences is deleted and the current infrared signal sequence is inserted at the end of the set X of infrared signal sequences. bg bg When no abnormal signal is detected in step S5.2, then the infrared signal sequence is updated in a way that the first infrared signal sequence in the set X of infrared signal sequences is deleted and the current infrared signal sequence is inserted at the end of the set X of infrared signal sequences.​

Images