How to track people in masks or a universal approach to tracking objects of arbitrary nature

Introduction

Since neural networks began to gain popularity, most engineers began to solve many of the problems of software in the field of Public Safety using deep learning methods. Despite the fact that neural networks have no competitors in terms of detection and identification of objects, they still cannot boast of the ability to analyze and reason, but only create patterns that cannot always be understood or interpreted.

We are of the opinion that interpretable and predictable approaches, such as, for example, the probabilistic data association approach, will be more effective for tracking multiple objects.

The tracking accuracy and the advantages of the approach we have chosen are clearly visible (more in the post):

Comparison of the popular Re3 tracker (left) and our component AcurusTrack (right)

: “ Public Safety Software”. , - . , .

GitHub.

Public Safety , . ( , , ..). , , .

, , ( "" ), , , , ..

, . , , .

, :

• , (, , dlib);
• , (, );
• , — ( ).

, , .

, ( ) , , . — . , , , .

N , . F_1, ..., F_N. (detections) D_1,...,D_k, k<=N. i=1,...,N, (id).

, , , . , , iou — intersection over union — k k-1 ; . , , , .

, , , , . , : OY y , O — x, , . , (x, y), i- , (i∗img_w+x, y), img_w — .

, ( OX )

, , EvenVizion, .

. , (data association).

, , - (MCMC).

() . , .

:

• , , ;
• , , , , . , ( );
• , ( : , );
• ;
• cpu .

, :

1. ;
2. EvenVizion;
3. ;
4. , .

AcurusTrack

, :

• ,

AcurusTrack , :

• : , ;
• .

computer vision. , , .. .

AcurusTrack , , .

, , ( ).

, — GitHub.