Lightweight fire smoke detection method, terminal device, and storage medium
The present invention pertains to the technical field of smoke detection and discloses a lightweight fire smoke detection method, a terminal device, and a storage medium. Primarily, a smoke-like foreground is screened out based on a combination of a mixed Gaussian model and a YUV color model; and an ACON activation function is introduced to replace a Leaky ReLU activation function of YOLOv5 to form an ACON-CSP module for feature extraction. In the present invention, the smoke foreground is extracted for preprocessing by using the combination of the mixed Gaussian and YUV. Thus, in the preprocessing stage, static interference and non-smoke color interference in the preprocessing stage are ruled out while narrowing down the range of smoke detection, ensuring the relative accuracy and improving the detection speed, thereby providing a possible method for fire smoke detection of a low-end terminal device in an outdoor computer.
1 . A lightweight fire smoke detection method, comprising the following steps:
obtaining a video image;
screening out a smoke-like foreground based on a combination of a mixed Gaussian model and a YUV color model;
calling an improved yolo algorithm to further screen the smoke-like foreground; and
predicting a specific position and range of smoke in each frame of image to complete a smoke detection process;
wherein the screening out a smoke-like foreground based on a combination of a mixed Gaussian model and a YUV color model specifically comprises:
performing pixel matching on each frame of video by using a mixed Gauss algorithm, to obtain a corresponding static background and dynamic foreground; and
determining an area growth rate of the dynamic foreground, and if the area growth rate is greater than 0, extracting the dynamic foreground again based on the YUV color model, to obtain a smoke foreground obtained via coarse screening to finally demarcate a specific position and range of smoke in the video;
wherein during the performing pixel matching on each frame of video by using a mixed Gauss algorithm, a weighted sum formula of a Gaussian distribution probability density function is:
P
(
x
t
)
=
∑
i
=
1
k
w
i
,
t
·
η
(
x
t
,
μ
i
,
t
,
τ
i
,
t
)
η
(
x
t
,
μ
i
,
t
,
τ
i
,
t
)
=
1
❘
"\[LeftBracketingBar]"
τ
i
,
t
❘
"\[RightBracketingBar]"
1
/
2
e
-
1
2
(
x
t
-
μ
i
,
t
)
T
τ
-
1
,
(
x
t
-
μ
i
,
t
)
τ
i
,
t
=
σ
i
,
t
2
I
(
1
-
1
)
wherein x t is a pixel in a current frame, k is the total number of Gaussian models, w i,t is weight of the i th Gaussian distribution at a moment t, η is a probability density function of the i th Gaussian distribution at the moment t, μ i,t , and τ i,t , are a mean and a covariance matrix, σ i,t 2 is a variance, and I is a three-dimensional identity matrix;
matching pixels in the current frame with preset mixed Gaussian models one by one until the Gaussian model that suits the pixel is found, as expressed in the following formula:
| x t -μ i,t |≤D·σ i,t (1-2)
wherein D is a confidence parameter and σ i,t , is a standard deviation of the i th Gaussian distribution at the moment t−1; and
if a Gaussian distribution matches the pixel successfully, setting M i,t to 1, and updating a parameter in the Gaussian model, as expressed in the following formula:
w i,t =(1-α) w i,t−1 +αM i,t
μ i,t =1−ρ)μ i,t−1 +ρx t
σ i,t 2 =(1−ρ)σ i,t−1 2 +ρ( x t −μ i,t ) T ( x t −μ i,t ) (1_3)
wherein α is a learning rate and ρ=α·η is a parameter update rate; or
if the Gaussian distribution does not match the pixel, keeping the mean μ i,t and variance σ i,t 2 in Gaussian distribution parameters unchanged, and only changing the weight w i,t , wherein M i,t is 0 in this case; and when none of k Gaussian distributions matches the pixel, creating a new Gaussian distribution to replace a Gaussian distribution with smaller weight in the original distributions, and performing normalization on the parameters; and
matching x t with the obtained B Gaussian models, and when one model satisfies a need of the pixel, treating the pixel as the background, otherwise, treating the pixel as the foreground, as expressed in the following formula:
B
=
arg
min
b
(
∑
i
=
1
b
w
i
,
t
>
T
)
(
1
-
4
)
wherein b is an initial parameter, and T is a preset threshold.
2 . The fire smoke detection method according to claim 1 , wherein during the extracting the dynamic foreground again based on the YUV color model, a formula for switching from RGB color space to YUV color space is as follows:
Y= 0.299 R+ 0.587 G+ 0.114 B
U=− 0.1687 R− 0.3313 G+ 0.5 B+ 128
V= 0.5 R− 0.4187 G− 0.813 B+ 128 (1-6)
wherein value ranges of Y and chrominance components U and V are all 0 to 255; and
a region that satisfies the following formula is determined as a smoke-like region:
I
color
=
{
I
(
x
,
y
)
U
(
x
,
y
)
>
T
U
OR
❘
"\[LeftBracketingBar]"
U
(
x
,
y
)
-
V
(
x
,
y
)
❘
"\[RightBracketingBar]"
>
T
UV
0
Others
(
1
-
7
)
wherein U(x,y) and V(x,y) are a component U and a component V of a point (x,y) respectively.
3 . A lightweight fire smoke detection method, comprising:
obtaining a video image;
screening out a smoke-like foreground based on a combination of a mixed Gaussian model and a YUV color model;
calling an improved yolo algorithm to further screen the smoke-like foreground; and
predicting a specific position and range of smoke in each frame of image to complete a smoke detection process;
wherein the calling an improved yolo algorithm to further screen the smoke-like foreground specifically comprises:
during extraction of a smoke feature, using a preprocessed smoke data set as input of the model, repeatedly training and adjusting a parameter in a mobile-yolov5 model, to finally generate a weight characterizing the smoke feature, and calling the weight for the smoke foreground obtained by coarse screening, to perform secondary fine screening on the smoke region;
wherein in the mobile-yolov5 model, a backbone network of YOLOv5s is replaced with a backbone network of Mobilenetv3, and an SE attention mechanism is added to the backbone network of mobile-yolov5;
introducing an ACON activation function to replace a Leaky ReLU activation function of YOLOv5 to form an ACON-CSP module for feature extraction.
4 . A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein when the processor executes the computer program, the method according to claim 1 is implemented.
5 . A non-transitory computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to claim 1 is implemented.