Method and apparatus for obstacle detection under complex weather

1 . A method for obstacle detection under complex weather, comprising:

obtaining an image under a complex weather condition;

performing enhanced preprocessing on the image by using a multi-scale retinex with color restoration MSRCR algorithm to obtain a preprocessed image;

inputting the preprocessed image into a trained obstacle detection model based on an improved YOLOv3 network; and

according to output of the obstacle detection model based on the improved YOLOv3 network, determining an obstacle detection result under the complex weather;

wherein a method for establishing the obstacle detection model based on the improved YOLOv3 network comprises:

inserting an SPP module into a convolutional set of YOLOv3 to fuse different scale characteristics for extraction of more detailed information to further improve detection accuracy, and replacing a Leaky-ReLU activation function in convolutional layers in the original YOLOv3 network with an ELU activation function to increase robustness to noise;

obtaining a data set for obstacle detection under different weather conditions to improve robustness of the training model;

re-clustering ground truth boxes in the data set by using a K-Means++ algorithm to obtain suitable anchor boxes, increasing a convergence speed of sample training and improving positioning accuracy; and

training the obstacle detection model with the processed data set to obtain the trained obstacle detection model based on the improved YOLOv3 network;

wherein the obtaining a data set for obstacle detection under different weather conditions to improve robustness of the training model comprises: selecting a representative KITTI data set and CODA data set, and supplementing the KITTI data set using the CODA data set with corresponding tags modified, to obtain the data set for obstacle detection in different weather conditions; and reserving a test set from the obtained data set, and using the rest of the data set as a training set to increase the robustness of the training model.

2 . The method for obstacle detection under complex weather according to claim 1 , wherein the performing enhanced preprocessing on the image by using a multi-scale retinex with color restoration MSRCR algorithm comprises:

employing a weighted sum of several different Gaussian scale parameters, using a Gaussian filtering result as an estimated illumination image, and adding a color restoration factor to restore color of a reflection image to avoid color degradation, with a calculation formula as follows:

F

MSRCR

(

x

,

y

)

=

C

i

⁢

∑

j

=

1

N

ω

j

[

lgI

i

(

x

,

y

)

-

lg

⁡

(

I

i

(

x

,

y

)

⊗

G

j

(

x

,

y

)

)

]

⁢

G

j

(

x

,

y

)

=

1

2

⁢

πσ

j

2

⁢

exp

⁡

(

-

x

2

+

y

2

2

⁢

σ

j

2

)

⁢

C

i

=

β

⁢

lg

[

α

⁢

I

i

(

x

,

y

)

∑

i

=

1

3

I

i

(

x

,

y

)

]

wherein F MSRCR (x, y) is an enhanced result of the processing by the MSRCR algorithm, x and y represent abscissa and ordinate of image pixels respectively, i represents an ordinal position number of a color component, j represents an ordinal position number of a scale parameter, C i is a color restoration factor of an i th color component, N is the number of scale parameters, ω j is weight corresponding to different scales, I i (x, y) represents an i th color component of the input image, G j (x, y) represents a Gaussian filter with a scale of σ j , σ j represents a scale parameter of a j th Gaussian surround, ⊗ is a convolution symbol, α is a nonlinear strength adjustment parameter, and β is a gain factor.

3 . The method for obstacle detection under complex weather according to claim 2 , wherein N=3, representing three scales: small, medium and large.

4 . The method for obstacle detection under complex weather according to claim 1 , wherein the SPP module comprises a total of four branches: a first branch is a branch with direct connection between input and output, a second branch is the maximum pooling with a 5×5 pooling core, a third branch is the maximum pooling with a 9×9 pooling core, and a fourth branch is the maximum pooling with a 13×13 pooling core, wherein a step size of each pooling is 1, and padding is performed before pooling to keep size and depth of a final characteristic map unchanged, so that the SPP module finally fuses different scale characteristics.

5 . The method for obstacle detection under complex weather according to claim 1 , wherein the replacing a Leaky-ReLU activation function in convolutional layers in the original YOLOv3 network with an ELU activation function comprises:

the ELU activation function being represented by a calculation formula:

u

=

{

e

m

-

1

,

m

≤

0

m

,

m

>

0

wherein u and m represent abscissa and ordinate of the function respectively, the ELU activation function has a non-zero gradient for all negative values, eliminating the problem of neuron death, wherein when abnormal input occurs in use of the ELU activation function, no large gradient is produced during backward propagation, preventing neuron death and vanishing gradient, which can shorten training time and improve accuracy in network training; when m in ELU is 0 or negative, an exponential function is used, and as the parameter decreases, the function gradually converges to a negative value; convergence means that there is a small derivative value, which reduces variation and information propagated to a next layer; and

therefore, the ELU activation function is more robust to noise, able to reduce impact of image noise on the detection result.

6 . The method for obstacle detection under complex weather according to claim 1 , wherein the supplementing the KITTI data set using the CODA data set with corresponding tags modified comprises:

categorizing obstacle tags into car, van, truck, pedestrian, pedestrian (sitting), cyclist, tram, and misc for detection and classification;

expanding a complex weather data set part of the CODA data set through Mixup, Cutmix, and Cutout data enhancement methods, and modifying corresponding tags to allow for conformance to the classification of obstacle tags; and

adjusting images in the data set to a uniform size of 416×416 pixels for ease of training of the obstacle detection model based on the improved YOLOv3 network.

7 . The method for obstacle detection under complex weather according to claim 1 , wherein the training the obstacle detection model with the processed data set comprises:

training the improved YOLOv3 network under the training set part of the data set, adjusting parameters according to an actual situation, and evaluating performance of the model by using mAP, a mean of various categories of AP, wherein a calculation formula is as follows:

mAP

=

∑

a

=

1

k

AP

a

k

wherein mAP refers to a mean of various categories of AP, AP α is an area under a precision-recall curve of the a th category, k represents the total number of categories, and a represents an ordinal position number of the category; and a larger value of the mean average precision mAP indicates a higher overall recognition accuracy of the model, with the reserved test set used to verify the effect, and the trained obstacle detection model based on the improved YOLOv3 network is thus obtained.

8 . An apparatus for obstacle detection under complex weather, comprising a processor and a non-transitory storage medium, wherein

the non-transitory storage medium is configured to store an instruction; and

the processor is configured to operate according to the instruction to perform the steps of the method of claim 1 .

9 . A non-transitory storage medium, storing a computer program, wherein when the computer program is executed by a processor, the steps of the method of claim 1 are implemented.