[paper note] Human Re-identification in Crowd Videos using Personal, Social and Environmental (PSE)

• ECCV 2016
• paper
• poster

Highlight

• Re-identification in crowded scenes. At the edge of human re-id in videos, dense crowd analysis and social force models.
• Model person, social, environment constraints.
  
  • Person: appearance and speed
  • Social: grouping, collision avoidance
  • Environment: transition probabilities between gates
• Experiment result is very bad on PRID 2011.

Constraints

• Appearance similarity
  
  • Relu6 and Fc7 layers of CNN features (VGG-19)
• Preferred speed
  
  • No rectify needed.
  • Assuming walking speed in a camera follow a normal distribution
• Destination
  
  • Captures the common and frequent patterns of movement between gates in different cameras
• Spatial grouping
  
  • Neerby people should walk through similar distance
• Social grouping
  
  • Reward individuals in the same group
• Collision Avoidance