(19)
(11)EP 3 683 540 A1

(12)EUROPEAN PATENT APPLICATION
published in accordance with Art. 153(4) EPC

(43)Date of publication:
22.07.2020 Bulletin 2020/30

(21)Application number: 19870064.3

(22)Date of filing:  01.03.2019
(51)International Patent Classification (IPC): 
G01C 1/00(2006.01)
G06T 7/00(2017.01)
G01C 11/04(2006.01)
G06N 3/08(2006.01)
(86)International application number:
PCT/CN2019/076719
(87)International publication number:
WO 2020/093631 (14.05.2020 Gazette  2020/20)
(84)Designated Contracting States:
AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR
Designated Extension States:
BA ME
Designated Validation States:
KH MA MD TN

(30)Priority: 06.11.2018 CN 201811317915

(71)Applicant: Wuyi University
Jiangmen, Guangdong 529000 (CN)

(72)Inventors:
  • WU, Yueting
    Jiangmen Guangdong 529000 (CN)
  • ZHAI, Yikui
    Jiangmen Guangdong 529000 (CN)
  • ZHENG, Yu
    Jiangmen Guangdong 529000 (CN)
  • ZHOU, Jihua
    Jiangmen Guangdong 529000 (CN)
  • WANG, Tianlei
    Jiangmen Guangdong 529000 (CN)
  • XU, Ying
    Jiangmen Guangdong 529000 (CN)
  • GAN, Junying
    Jiangmen Guangdong 529000 (CN)
  • DENG, Wenbo
    Jiangmen Guangdong 529000 (CN)
  • KE, Qirui
    Jiangmen Guangdong 529000 (CN)

(74)Representative: Juncosa Miró, Jaime et al
Torner, Juncosa i Associats, S.L. Gran Via de les Corts Catalanes, 669 bis, 1º, 2ª
08013 Barcelona
08013 Barcelona (ES)

  


(54)ANTENNA DOWNTILT ANGLE MEASUREMENT METHOD BASED ON DEPTH INSTANCE SEGMENTATION NETWORK


(57) A method for measuring an antenna downtilt angle based on a deep instance segmentation network is disclosed, including: shooting an omni-directional antenna video using a drone; and transmitting the antenna video to a server in real time, and the server measuring an antenna downtilt angle in real time using a deep learning algorithm, the deep learning algorithm includes: a feature extraction network module for acquiring an antenna feature image; an instance segmentation module for binary segmentation of an antenna image and the background to distinguish antenna image pixels from background pixels; an antenna candidate box module for identifying and detecting the antenna image, acquiring an antenna candidate box and determining an antenna calibration box therefrom; and an antenna downtilt angle measuring module for measuring an antenna downtilt angle; and the instance segmentation module includes: directly mapping a region of interest to a feature map; segmenting a candidate region into kk cells, determining, for each cell, four fixed coordinate positions of the center point of the cell, and calculating values of the four fixed coordinate positions by bilinear interpolation; and performing max-pooling and implementing back propagation. By distinguishing antenna pixels from background pixels in the instance segmentation module, the antenna downtilt angle is more accurate, and a convenient, safe, effective and accurate antenna measuring method is established.




Description

TECHNICAL FIELD



[0001] The present disclosure relates to the field of communication measurement, and in particular, to a method for measuring an antenna downtilt angle based on a deep instance segmentation network.

BACKGROUND



[0002] An antenna downtilt angle often needs to be adjusted in the field of communications. The antenna downtilt angle is one of the important parameters to determine the coverage of base station signals. It is not only necessary to accurately design a downtilt angle of each antenna at the initial stage of network planning, but also necessary to accurately adjust the downtilt angle after a base station is put into operation with the development of services, changes of users and changes of an ambient signal environment.

[0003] At present, a mechanical downtilt angle of a base station antenna is generally measured by a gradiometer. When measuring the mechanical downtilt angle of the antenna with the gradiometer, a surveyor must climb an iron tower or hold a pole close to the antenna for measurement, which is quite dangerous and troublesome, and also affects the accuracy of measurement. With the technological development, a GSM-R system emerges. The system is a measuring system tool with which a surveyor can accurately measure an antenna downtilt angle without getting close to an antenna, which can realize the measurement of a tilt angle of a base station antenna without a tower climbing operation, and can network test points of each base station to realize the real-time monitoring of the downtilt angle of the base station antenna. However, the installation of a sensor is time-consuming and costly, and there are differences in old and new towers and the number of layers and the number of base station towers, so the method is not practical, has a long operation cycle, and is difficult to implement. Therefore, it is necessary to design an angle measuring method with simple operations and reliable performance.

SUMMARY



[0004] To solve the above problems, the embodiments of the present disclosure aim at providing an antenna downtilt angle measuring method based on a deep instance segmentation network, so as to measure an antenna downtilt angle safely, effectively, quickly and accurately.

[0005] The technical scheme adopted by the embodiments of the present disclosure to solve the problems is as follows:
An antenna downtilt angle measuring method based on a deep instance segmentation network, comprising: shooting an omni-directional antenna video using a drone; and transmitting the antenna video to a server in real time, and the server measuring an antenna downtilt angle in real time using a deep learning algorithm wherein the deep learning algorithm comprises: a feature extraction network module for acquiring an antenna feature image; an instance segmentation module for binary segmentation of an antenna image and the background to distinguish antenna image pixels from background pixels; an antenna candidate box module for identifying and detecting the antenna image, acquiring an antenna candidate box and determining an antenna calibration box therefrom; and an antenna downtilt angle measuring module for measuring an antenna downtilt angle; and the instance segmentation module comprises: directly mapping a region of interest to a feature map; segmenting a candidate region into kk cells, determining, for each cell, four fixed coordinate positions of the center point of the cell, and calculating values of the four fixed coordinate positions by bilinear interpolation; and performing max-pooling and implementing back propagation.

[0006] Further, the shooting an omni-directional antenna video using a drone comprises: controlling the drone to fly around in a circle with an antenna as the origin and a radius of r meters from the antenna at a height of h meters from the antenna, and shooting a 360-degree omni-directional antenna video.

[0007] Further, the feature extraction network module comprises a residual network for solving gradient problems and network performance degradation problems and a feature pyramid network for solving multi-scale detection problems.

[0008] Further, a back propagation formula of the instance segmentation module is

wherein xi represents a pixel point on the feature map before max-pooling; yrj represents the jth point in the rth candidate region after max-pooling; i(r, j) represents coordinates of the pixel point before max-pooling corresponding to a pixel value of the point yrj; xi(r, j) represents coordinates of sampling points calculated during back propagation, d(.) denotes a distance between two points, and Δh and Δw denote differences of horizontal and vertical coordinates of xi and xi(r, j).

[0009] Preferably, the antenna candidate box module is a candidate box selection module of a faster RCNN network structure.

[0010] Further, a target loss function of the antenna candidate box module is

wherein Lcls is a classification loss function, and its expression is:

Lreg is a regression loss function, and its expression is:

and an R function is defined as Smooth



[0011] Further, the antenna downtilt angle measuring module obtains a region with a minimum relative proportion by comparing a relative proportion between the antenna calibration box and a region represented by the antenna image pixels.

[0012] Preferably, the antenna downtilt angle measuring module measures the antenna downtilt angle based on the following formula:

wherein x is the width of the calibration box, and y is the length of the calibration box.

[0013] The embodiments of the present disclosure have the following beneficial effects: the embodiments of the present disclosure adopt an antenna downtilt angle measuring method based on a deep instance segmentation network, in which an antenna image to be measured is inputted into a deep learning network and is sequentially processed by a feature network module, an instance segmentation module, an antenna candidate box module and an antenna downtilt angle measuring module to output an antenna downtilt angle. Meanwhile, the instance segmentation network achieves binary segmentation of antenna pixels and background pixels, which makes the subsequent antenna candidate region more simple and accurate, so that the measured antenna downtilt angle is more consistent with an actual value. According to the embodiments of the present disclosure, the danger of climbing measurement is avoided and the cost of installing the sensor is reduced, and data of the antenna downtilt angle can be obtained more effectively, safely, cheaply and accurately.

BRIEF DESCRIPTION OF THE DRAWINGS



[0014] The present disclosure is further described below with reference to the accompanying drawings and examples.

Fig. 1 is a network structure diagram of a deep learning algorithm according to an embodiment of the present disclosure;

Fig. 2 is a structural diagram of a residual network of a feature extraction network module according to an embodiment of the present disclosure;

Fig. 3 is a structural diagram of a feature pyramid network of a feature extraction network module according to an embodiment of the present disclosure;

Fig. 4 is a schematic flowchart of an instance segmentation module according to an embodiment of the present disclosure; and

Fig. 5 is a structural diagram of an antenna candidate box module according to an embodiment of the present disclosure.


DETAILED DESCRIPTION



[0015] Referring to Fig. 1, an antenna downtilt angle measuring method based on a deep instance segmentation network is disclosed in an embodiment of the present disclosure, including: shooting an omni-directional antenna video using a drone; and transmitting the antenna video to a server in real time, and the server measuring an antenna downtilt angle in real time using a deep learning algorithm, wherein the deep learning algorithm includes: a feature extraction network module for acquiring an antenna feature image; an instance segmentation module for binary segmentation of an antenna image and the background to distinguish antenna image pixels from background pixels; an antenna candidate box module for identifying and detecting the antenna image, acquiring an antenna candidate box and determining an antenna calibration box therefrom; and an antenna downtilt angle measuring module for measuring an antenna downtilt angle.

[0016] In an embodiment, the shooting an omni-directional antenna video using a drone includes: controlling the drone to fly around in a circle with an antenna as the origin and a radius of r meters from the antenna at a height of h meters from the antenna, and shooting a 360-degree omni-directional antenna video, wherein h is 10, and r is 10.

[0017] In an embodiment, the feature extraction network module includes a residual network for solving gradient problems and network performance degradation problems and a feature pyramid network for solving multi-scale detection problems.

[0018] Referring to Fig. 2, in a deep learning network, the "level" of features will increase as the depth of the deep learning network deepens, so the depth of the deep learning network is an important factor to achieve a good feature extraction effect. However, the gradual deepening of the network depth will be accompanied by gradient disappearance and gradient explosion and other problems. The conventional solution is to initialize and regularize data, which solves the gradient problems but leads to problems of performance degradation and an increased error rate of the deep learning network. The use of the residual network can solve the gradient problems as well as the problem of performance degradation of the deep learning network.

[0019] Referring to Fig. 3, the feature pyramid network is designed with a top-down and horizontally connected structure to combine a shallow layer with high resolution and a deep layer with rich semantic information. The top-down structure first samples an antenna feature map in a deeper layer with stronger semantics, and then connects the sampled antenna feature horizontally to a feature layer in the previous layer, thus reinforcing the antenna feature in the higher layer. The feature pyramid network can quickly construct a feature pyramid with strong semantic information at all scales from a single-scale input antenna image without a significant cost. With simple network connection changes, the performance of antenna object detection is greatly improved without increasing the calculation amount of the original model.

[0020] Referring to Fig. 4, in an embodiment, the instance segmentation module includes: directly mapping a region of interest to a feature map; segmenting a candidate region into kk cells, determining, for each cell, four fixed coordinate positions of the center point of the cell, and calculating values of the four fixed coordinate positions by bilinear interpolation; and performing max-pooling and implementing back propagation.
a back propagation formula of the instance segmentation module is

wherein xi represents a pixel point on the feature map before max-pooling; yrj represents the jth point in the rth candidate region after max-pooling; i(r, j) represents coordinates of the pixel point before max-pooling corresponding to a pixel value of the point yrj; xi(r, j) represents coordinates of sampling points calculated during back propagation, d(.) denotes a distance between two points, and Δh and Δw denote differences of horizontal and vertical coordinates of xi and xi(r, j), which are multiplied by the original gradient as a bilinear interpolation coefficient. In the antenna feature map before max-pooling, each point and the horizontal and vertical coordinates of xi(r, j) which are less than 1 will receive a gradient returned by the point yrj corresponding thereto.

[0021] Referring to Fig. 5, in an embodiment, preferably, the antenna candidate box module is a candidate box selection module using a faster RCNN network structure.

[0022] Further, the antenna candidate box module finally classifies and regresses the selected candidate box, its target loss function is

wherein Lcls is a classification loss function, and its expression is:

Lreg is a regression loss function, and its expression is:

and an R function is defined as Smooth



[0023] Further, the antenna downtilt angle measuring module obtains a region with a minimum relative proportion by comparing a relative proportion between the antenna calibration box and a region represented by the antenna image pixels.

[0024] Preferably, the antenna downtilt angle measuring module measures the antenna downtilt angle based on the following formula:

where x is the width of the calibration box, and y is the length of the calibration box.

[0025] The above are merely preferred embodiments of the present disclosure, and the present disclosure is not limited to the above implementations. All implementations shall be encompassed in the protection scope of the present disclosure as long as they achieve the technical effect of the present disclosure with the same means.


Claims

1. A method for measuring an antenna downtilt angle based on a deep instance segmentation network, comprising:

shooting an omni-directional antenna video using a drone; and

transmitting the antenna video to a server in real time, and the server measuring an antenna downtilt angle in real time using a deep learning algorithm,

wherein the deep learning algorithm comprises: a feature extraction network module for acquiring an antenna feature image; an instance segmentation module for binary segmentation of an antenna image and the background to distinguish antenna image pixels from background pixels; an antenna candidate box module for identifying and detecting the antenna image, acquiring an antenna candidate box and determining an antenna calibration box therefrom; and an antenna downtilt angle measuring module for measuring an antenna downtilt angle; and

the instance segmentation module comprises: directly mapping a region of interest to a feature map; segmenting a candidate region into kk cells, determining, for each cell, four fixed coordinate positions of the center point of the cell, and calculating values of the four fixed coordinate positions using bilinear interpolation; and performing max-pooling and implementing back propagation.


 
2. The method for measuring an antenna downtilt angle based on a deep instance segmentation network according to claim 1, wherein the shooting an omni-directional antenna video using a drone comprises: controlling the drone to fly around in a circle with an antenna as the origin and a radius of r meters from the antenna at a height of h meters from the antenna, and shooting a 360-degree omni-directional antenna video.
 
3. The method for measuring an antenna downtilt angle based on a deep instance segmentation network according to claim 2, wherein the feature extraction network module comprises a residual network for solving gradient problems and network performance degradation problems and a feature pyramid network for solving multi-scale detection problems.
 
4. The method for measuring an antenna downtilt angle based on a deep instance segmentation network according to claim 1, wherein the back propagation formula of the instance segmentation module is

wherein xi represents a pixel point on the feature map before max-pooling; yrj represents the jth point in the rth candidate region after max-pooling; i(r, j) represents coordinates of the pixel point before max-pooling corresponding to a pixel value of the point yrj; xi(r, j) represents coordinates of sampling points calculated during back propagation, d(.) denotes a distance between two points, and Δh and Δw denote differences of horizontal and vertical coordinates of xi and xi(r, j).
 
5. The method for measuring an antenna downtilt angle based on a deep instance segmentation network according to claim 1, wherein the antenna candidate box module is a candidate box selection module using a faster RCNN network structure.
 
6. The method for measuring an antenna downtilt angle based on a deep instance segmentation network according to claim 5, wherein the antenna candidate box module's target loss function is

wherein Lcls is a classification loss function, and its expression is:



Lreg is a regression loss function, and its expression is:

and

an R function is defined as Smooth


 
7. The method for measuring an antenna downtilt angle based on a deep instance segmentation network according to claim 1, wherein the antenna downtilt angle measuring module obtains a region with a minimum relative proportion by comparing a relative proportion between the antenna calibration box and a region represented by the antenna image pixels.
 
8. The method for measuring an antenna downtilt angle based on a deep instance segmentation network according to claim 7, wherein the antenna downtilt angle measuring module measures the antenna downtilt angle based on the following formula:

where x is the width of the calibration box, and y is the length of the calibration box.
 




Drawing
















Search report