Eureka AIR delivers breakthrough ideas for toughest innovation challenges, trusted by R&D personnel around the world.

A Dense Population Counting and Localization Method Based on Distance Transformed Labels

A technology of distance transformation and positioning method, which is applied in the field of computer vision, can solve the problems of inability to provide positioning information and inapplicable to dense crowd scenes, etc., and achieve the effect of simple counting method

Active Publication Date: 2022-04-26
HUAZHONG UNIV OF SCI & TECH
View PDF4 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0004] The present invention proposes a dense group counting and positioning method based on distance transformation labels, which can provide specific location information of each head while ensuring counting accuracy, so as to solve the problem that the method based on regression density map in dense group scenes cannot provide Accurate positioning information, the problem that detection-based methods cannot be applied to dense crowd scenes

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • A Dense Population Counting and Localization Method Based on Distance Transformed Labels
  • A Dense Population Counting and Localization Method Based on Distance Transformed Labels
  • A Dense Population Counting and Localization Method Based on Distance Transformed Labels

Examples

Experimental program
Comparison scheme
Effect test

example 1

[0069] (1) Make a data set. For each head, due to the innovation of this method, only the center coordinates of each head need to be provided, and the labeling frame of each head is not required. Compared with the detection-based method, the cost of human labeling is greatly reduced. Online expansion of the produced data set, the specific operations are random horizontal flip, random scaling between 0.8-1.5 times, and random superposition of salt and pepper noise. The number of noise points of salt and pepper noise is 0.001-0.015 times the number of pixels in the picture.

[0070] (2) Generate a binary map Binary map. Pre-set the two-dimensional zero matrix. For a given center coordinate (x, y) of each head, according to the coordinate position, use 1 to represent the point on the matrix, such as figure 1 Shown represents the original image, figure 2 Shows the Binary map generated from the coordinates of the center of each head. It should be pointed out that, for intuitiv...

example 2

[0103] According to the method proposed in the present invention, the performance obtained on three common public datasets is given, specifically including ShanghaiTech_partA, ShanghaiTech_partB and UCF-QNRF.

[0104] The ShanghaiTech_partA dataset contains 482 pictures, including 300 training pictures and 182 test pictures, which are mainly crawled from the Internet. The density of the data set is relatively moderate, and the average number of people is 501.

[0105] The ShanghaiTech_partB data set contains 716 pictures, of which 400 are used for training and 316 are used for testing. The data set is mainly from the streets of Shanghai. The density of the data set is relatively sparse, and the average number of people is 123.

[0106] UCF-QNRF contains 1535 pictures, of which 1201 are used for training and 334 are used for testing. The data set is very dense. The average number of people is 815, and the maximum number reaches 12865.

[0107] The evaluation indicators used are...

example 3

[0121] scalability. This method can not only be used to complete the counting and positioning of crowds, but also can be used to complete the counting and positioning of other objects, such as vehicles and cells. Here, the present invention takes the public vehicle data set trancos as an example to illustrate its ability in vehicle counting and positioning.

[0122] First, the generated Distance label map is also used as the real label D. Since the dataset provides mask annotations, that is, the region of interest, the input image point is multiplied by the mask as input, and the output point is also multiplied by the mask. Using the same FPN structure as crowd counting and positioning, the loss function remains unchanged, and the test results after training are shown in Figure 9 , where GT represents the number of vehicles in the ground truth label, and Count represents the number of predicted vehicles, showing excellent counting and localization accuracy.

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention discloses a dense group counting and positioning method based on distance transformation labels. The present invention proposes a novel real label generation method, which is called Distance label map. Through the Distance label map, it is only necessary to count the number of local minimum value regions to obtain the estimated number of people. In addition, calculate The center coordinates of the local minimum area can obtain the specific position information of the predicted head. The feature extraction adopts the feature pyramid network, and the picture is input into the network to obtain multiple different fusion features, and the loss is calculated and summed with the real label. This invention proposes a new function for calculating the loss, which is called adaptive cross entropy , can use the label of the pixel point to represent the distance information of the pixel point from its nearest pixel point to carry out weighted improvement on the traditional cross entropy. Compared with the previous work related to group counting, the present invention can provide positioning information of each head while ensuring counting accuracy.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a dense group counting and positioning method based on distance transformation labels. Background technique [0002] In recent years, stampede incidents frequently occur when large-scale events are held at home and abroad. The main reason is that the specific number and location of the crowd cannot be known in time, and the flow of people cannot be effectively evacuated. This makes the crowd counting work in video surveillance, crowd understanding, and public safety. Prevention and other fields have gradually played an important role, and have become hot research issues in the field of computer vision. [0003] With the development of deep learning, researchers can use deep neural networks to achieve dense crowd counting. The existing crowd counting methods can be generally divided into two categories. The first is to use crowd counting as a target detection task, and pedestrians o...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Patents(China)
IPC IPC(8): G06V20/52G06K9/62G06V10/774G06V20/40
CPCG06V20/41G06V20/53G06F18/214
Inventor 许永超徐晨丰梁定康白翔
Owner HUAZHONG UNIV OF SCI & TECH
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Eureka Blog
Learn More
PatSnap group products