Computational Social Cognition: Perception, Modeling, and ...hspark/GroupBehavior/...Choi, ECCV12,...
Transcript of Computational Social Cognition: Perception, Modeling, and ...hspark/GroupBehavior/...Choi, ECCV12,...
Social Dynamics
Mathematical Definition of Group Behaviors
Behaviors of X and Y are a group behavior if spatial/temporal predictions of Y based on X and Y jointly are more accurate than predictions based on Y alone.
Y X
ˆ ˆ∆ − ∆ ≥ ∆ − ∆ Y(t + t) Y(t + t | Y(0), ,Y(t)) Y(t + t) Y(t + t | Y(0), ,Y(t),X(0), ,X(t))
Inspired by Granger Causality
Number of group members
Scen
e dy
nam
ism
Dyadic interaction Crowd interaction
Static
scen
e Dy
nami
c sce
ne
Rehg, CVPR13 Prabhaker, ECCV12 Prabhakar, CVPR12 Patron-Perez, BMVC10
Yang, CVPR12 Hoai, CVPR14
Fathi, CVPR12 Choi, ECCV14 Park, NIPS12, ICCV13
Rodriguez, ICCV11a, ICCVb Mehran, CVPR09 Alahi, CVPR14
Cristani, BMVC11 Park, CVPR15 Arev, SIGGRAPH14
Lan, CVPR12 Ramanathan, CVPR13 Antic, ECCV14
Wang, ECCV10 Gallagher, CVPR09
Ding, ECCV10 Choi, ECCV12, CVPR14 Direkoglu, ECCV12
Number of group members
Scen
e dy
nam
ism
Dyadic interaction Crowd interaction
Static
scen
e Dy
nami
c sce
ne
Rehg, CVPR13 Prabhaker, ECCV12 Prabhakar, CVPR12 Patron-Perez, BMVC10
Yang, CVPR12 Hoai, CVPR14
Fathi, CVPR12 Choi, ECCV14 Park, NIPS12, ICCV13
Rodriguez, ICCV11a, ICCVb Mehran, CVPR09 Alahi, CVPR14
Cristani, BMVC11 Park, CVPR15 Arev, SIGGRAPH14
Lan, CVPR12 Ramanathan, CVPR13 Antic, ECCV14
Wang, ECCV10 Gallagher, CVPR09
Ding, ECCV10 Choi, ECCV12, CVPR14 Direkoglu, ECCV12
Measurements
Individual social signal
Group behavior
Social role
Social Statics
Measurements
Individual social signal
Group behavior
Social role
Time
Social Dynamics
Measurements
Individual social signal
Group behavior
Social role
Social Dynamics
Model based Approach
Measurements
Individual social signal
Group behavior
Social role
Causal Relationship
Input: a video of social interactions Output: to find a causal relationship
[Prabhakar, CVPR10]
Measurements
Individual social signal: motion
Group behavior: Causality
Social role
Time Signal Representation Point Process
Point process A random process for a set of isolated points either in time or space.
events
1 1 2 2 2 3 3 4 4 5 6 7 7 7 7 7 8 8 9 N(t)
0 1 2 3 3 3 4 5 5 7 7 7 8 9 10 11 11 13 13 N(t)
Multivariate point process
[Prabhakar, CVPR10]
Visual Signal Representation Point Process
[Prabhakar, CVPR10]
Causal Representation Cross Correlation
Cross correlation between point processes:
Spectral matrix for the multivariate point process: auto-spectrum+cross-spectrum
where
[Prabhakar, CVPR10]
Causal Relationship Detection
Green: causally related actions
[Prabhakar, CVPR10]
Causal Relationship Detection
[Prabhakar, CVPR10]
Causal Relationship Detection
[Prabhakar, CVPR10]
Social Force Model
Input: pedestrian locations Model: Newtonian mechanics
[Helbing, Phys. Rev. 95, Johansson, Adv. Comp. Sys. 2008]
Measurements
Individual social signal: position
Group behavior: Newtonian mechanics
Social role
Equation of Motion for Pedestrians
αrα
krα : destination
eα
vα : instantaneous velocity
Fα
1) He or she wants to reach a certain destination.
Acceleration
[Helbing, Phys. Rev. 95]
Equation of Motion for Pedestrians
αrα
krα : destination
eα
vα
Fα β
: Potential by wall Vαβ : Potential field by βwallU
i
1) He or she wants to reach a certain destination.
Acceleration
2) Motion of a pedestrian is influenced by other pedestrian. Repulsive effect induced by β
Deceleration induced by wall
Attractive effect induced by i
[Helbing, Phys. Rev. 95]
Equation of Motion for Pedestrians 1) He or she wants to reach a certain destination.
Acceleration
2) Motion of a pedestrian is influenced by other pedestrian.
αrα
krα : destination
eα
vα
Fα αβ
Repulsive effect induced by β
: Potential by wall Vαβ : Potential field by βwallUDeceleration induced by wall
i
Attractive effect induced by i
3) Line of sight effect
[Helbing, Phys. Rev. 95]
Equation of Motion for Pedestrians
Equation of motion:
αrα
krα : destination
eα
vα
Fα αβ
: Potential by wall Vαβ : Potential field by βwallU
i
ddt
=(preferred velocity)
(Force / acceleration) + (Noise)
[Helbing, Phys. Rev. 95]
wα
wα : preferred velocity affected by social force model.
Pedestrian Stress Analysis
Jamarat bridge
[Johansson, Adv. Comp. Sys. 2008]
Pedestrian Stress Analysis
Jamarat bridge
[Johansson, Adv. Comp. Sys. 2008]
Abnormal Crowd Behaviors
Input: pedestrian tracks Output: to detect abnormal behaviors
[Mehran, CVPR09]
Measurements
Individual social signal: crowd motion
Group behavior: Newtonian mechanics
Social role
Generalized Social Force Model
Social force model
Adjustment for desired velocity
Interaction between pedestrians
Generalized social force model
[0,1]ip ∈where is a panic parameter.
piv
civ
civ : average velocity of the neighboring pedestrians.
[Mehran, CVPR09]
Generalized Social Force Model
Interaction force Optical flow field
[Mehran, CVPR09]
[Mehran, CVPR09]
Crowd Activity Forecasting
Input: short term trajectories Output: long term trajectories
[Alahi, CVPR14]
Measurements
Individual social signal: motion
Group behavior: social affinity
Social role
100 m / 330 ft. 10 m 33 ft.
A Corridor with 14 Origin/Destination (O/D) Top View
O/D O/D O/D O/D O/D
O/D
O/D
O/D O/D O/D O/D O/D
O/D
O/D
Origin 5
Destination 11
100 m / 330 ft. 10 m 33 ft.
O/D O/D O/D O/D O/D
O/D
O/D
O/D O/D O/D O/D O/D
O/D
O/D
Origin 5
Destination 11
Study the performance of having few cameras to estimate the O/D matrices
A Corridor with 14 Origin/Destination (O/D) Top View
Collect long-term trajectories
31
# People Av. duration Av distance Density (up to) # Paths (O/D)
42 million 1 minute 100m 1 pedestrian/m2 196
Density
Collect long-term trajectories
32
# People Av. duration Av distance Density (up to) # Paths (O/D)
42 million 1 minute 100m 1 pedestrian/m2 196
Social Affinity Map (SAM)
Encoding spatial positions of neighbors
Learning Social Affinity Map
Heatmap of the most occupied regions surrounding an individual
Direction of motion
Input: • X: tracklets
Time t0 Time t1 Time tn
Tracking in a Large-scale Network
Tracking in a Large-scale Network
Input: • X: tracklets
Output: • T: all long term trajectories
Linked Tracklets
No prior
SAM prior
Model based Approach
Measurements
Individual social signal
Group behavior
Social role