1 System state model
1.1 Definition of state variables
bbox : the location of the target box;
(In the world of GoK, the speed of an object is 0 without external force)
( Class : the detection module will output the category of the target)
Remarks:
( Loc_txt : If the amount of information is not enough, you can also use txt location for auxiliary fusion positioning)
( Key : This modeling method is temporarily not used, and it feels that the granularity is fine, and the requirements on the model will be very high)
1.2 State transition equation
A : state transition matrix,
A = E ;
For other objects, external force factors need to be used for modeling, and the relative displacement of the target needs to be added;
2 Remarks
2.1 The fundamental solution of timing information fusion I think it is still LSTM
In my opinion, Kalman filtering is still a modeling method based on a priori information. If you want to implement a complete end-to-end model, you still need to use LSTM, a "learning" algorithm to fuse timing information;
2.1 What algorithm is used to predict the tracking module of Autoware?
Shi Fang said this:
2.2 You can also use CNN + LSTM to implement the tracking function
You can take a look at this article "Zero to Hero: A Quick Guide to Object Tracking: MDNET, GOTURN, ROLO" ,
It says something like this: