This is the design document of the Data association part of the object tracking node

Purpose / Use cases

The data association module pairs up detections with tracks that are similar so that the new state of the tracks can be better estimated or new tracks can be created or existing tracks can be pruned.

Design

The data association module uses Mahalanobis distance to compute how "similar" a track is to a detection. This cost is used to solve a hungarian assigner problem which returns a one to one association between detections and tracks. It is possible that some tracks/detection do not have any corresponding association.

Assumptions / Known limits

• Detections are assumed to be convex polygons with non zero area
• Detections are assumed to have their vertices ordered counterclockwise
• Each track in the tracks message has a vector of shapes to support articulated vehicles. The associator just uses the first shape in the vector for the purposes of association

Inputs / Outputs / API

Inputs:

• List of detections
• List of tracks

Parameters:

• Max area ratio between a track and its associated detection
• Max euclidean distance allowed between a track and its associated detection

Output

• Associations between track and detections
  

API:

• Call assign() with list of detection and tracks.
  This will return a struct that contains,
  
  1. Index of the associated detection for each track, or AssociatorResult::UNASSIGNED if a track could not be associated to any detection
     
  1. Indices of all the detections that do not have an associated track
     
  1. Indices of all the tracks that do not have an association detection

Inner-workings / Algorithms

• Figure out number of tracks and detections received and initialize the assigner appropriately
  
• Loop through detections and tracks. For each pair check if they can be associated based on the configured parameters
  
• If a pair meets the gating parameters, compute the Mahalanobis distance between them and assign a weight to the pair in the assigner. Otherwise, ignore that pair
  
• Call assign() function in the assigner
  
• Loop through all the associations to figure out the tracks and detections with no associations
  

Error detection and handling

• If a track or detection shape does not meet the assumptions mentioned above, exception is thrown

Future extensions / Unimplemented parts

• Consider many-to-one associations
  
• Consider geometric similarity metrics
  

Related issues

• #897 - Initial implementation