Ideas for Feature-basedTracking Improvements

• feature tracking
  • keep looking for optical flow matches for lost features (then interpolate linearly if found again)
  • keep stationary features who have moved enough in the past (remove stationary part if they do not move again)
  • detect keypoints continuously over the whole image to check that a feature trajectory would have some level of continuous detection
  • run shadow detection and remove features mostly detected as shadow over time
• avoid over grouping

stats on motion model: acceleration, speed for normal and overgrouped

• Classify features (based on speed, image patch, image foreground) and connect accordingly

compute gait/kinematic parameters and connect accordingly

• Add and modulate the similarity constraints

constant distance over time similar velocity angle robust: percentage of time that the condition must be met

• Extend the graph with weights instead of 0/1 (edge or not)