Tracking System

跟踪算法使用：
（1）每帧检测，一般不需要跟踪算法。 Detect with no track (No track. Need to smooth probably)
如果检测算法很快，则不需要进行跟踪，反之，嘿嘿
id 平滑
bbox 平滑，比如边框大小

（2）隔帧检测，需要跟踪。 Need to track

• 检测（当前帧进行检测） – A
  detection <==> observation
  We have detection this frame
  Detect 测量值

  Update with detection
  1）No instances (initial frames) —> Add detected into instances

  2）We have instances:
  a. Prediction (Kalman Filter)
  b. Match (prediction - Detection) —> Munkres Algorithm / Bipartite Algorithm or Double Match
  怎样预知匹配对应情况？

  3）Update newly matched instance

  4）Deal with failed matched instance
  4.1）Failed to detect
  a. Still track for a while
  b. Remove
  4.2）Newly detected
  a. Add instantly
  b. Add after a while （代码实现）

• 跟踪（此帧不检测） – B
  纯跟踪 prediction
  We have no detection this frame: Update without detection