Deep Sort

题目:SIMPLE ONLINE AND REALTIME TRACKING WITH A DEEP ASSOCIATION METRIC

作者: Nicolai Wojke †, Alex Bewley? Dietrich Paul †

 

 

1. sort based on the depth information is added: REID, IDswith 45% reduction;

2. Direct detection with neglect of the class distinction;

3.kalman: X=[Xcenter, Ycenter, ratio, Hight, Xcenter* , Ycenter*, ratio* ,Hight*]

4. The appearance characteristics stored and then compared,

The time allowed is 30;

6.C(i,j)=@*d1(i,j)+(1-@)*d2(i,j)

 d1: kalman amount and predicted observables

 The apparent features cosine distance: d2

7. Code: matching said cascade: REID → IOU, which contains the location of the threshold limit

8 is a front three experimental

 

problem found:

1. The association is not accurate, identity transformation occurs, but this is after reduced;

2.FP higher, as explained herein observation time is too long resulting 30;

3. Replace REID effect how bad it

 

Thinking:

1. only use location information, did not really use the speed information;

2. The optical flow information

 

Guess you like

Origin www.cnblogs.com/ciweizailianai/p/11314029.html