Title: Floor Field Models for Tracking in High Density Crowds
1Floor Field Models for Tracking in High Density
Crowds
- Saad Ali and Mubarak Shah
- Computer Vision Lab
- University of Central Florida
- ECCV 2008
2Tracking
Single Point
Multiple Points
Bounding Box
Contour
3Traditional Approaches to Tracking
- Features
- Appearance
- Color, texture
- Motion
- Shape
- Motion Correspondence
- Multipoint Correspondences
- Parametric transformation of a patch
- Contour Evolution
4Object Tracking A Survey
A. Yilmaz, O. Javed, and M. Shah, Object
Tracking A Survey, ACM Journal of Computing
Surveys, Vol. 38, No. 4, 2006.
5Crowded Scenes
6Homography Constraint
Occluded Feet detected!
H1to2
Plane parallax error
View 1
View 2
Feet Segmented
7Tracking Using Multiple Cameras
8Related Work - Detection in (of) Crowds
9High Density Crowded Scenes
High Density Moving Objects
Political Rallies
Religious Festivals
Marathons
10Related Work
- Only works in moderate density scenarios.
- Will fail in high density scenes where it is
difficult to establish ownership of features. - Approaches are object centric and do not exploit
scene and contextual knowledge.
11Tracking in High Density Crowds
12Challenges
- Tracking in high density scene is challenging
- Number of pixels on the object decreases with the
increase in the density. - Hard to discern individual from each other due to
constant interaction among them. - Inter-object occlusion.
- Dynamics of a crowd itself is complex
- Goal Directed
- Psychological characteristics
13Approach
- Observation
- Behavior of an individual in a crowded
situation is a function collective behavioral
patterns evolving from the space-time interaction
of a large number of individuals among
themselves, and with the structure of the scene
14Approach
- This collective behavioral Scene Structure can
be channeled in as an auxiliary source of
information. - Constrains likely locations and paths.
- Scene Structure Force Model
- in terms of Floor fields
- Inspired by evacuation dynamics
15Overview
- Treat the crowd flow as a collection of mutually
interacting particles. - Reasonable assumption, because when people are
densely packed, individual movement is restricted.
16Overview
- Model instantaneous movement of the individual
with a matrix of preferences. - Takes into consideration multiple sources of
information. - Appearance
- Scene Structure
17Overview
- Each cell encodes probability of move in a
certain direction. - Takes into consideration multiple sources of
information. - Appearance
- Scene Structure
18Scene Structure
- Encoded in terms of Floor Fields
- Floor Fields
- Model the interaction between individuals and
their preferred direction of movement by
transforming the long ranged forces into local
ones. - E.g. Long range force compelling the individual
to move towards the exit door can be converted
into a local force. - Probability of move of an individual depends on
the strength of the floor field in his/her
neighborhood.
19Floor Fields
Instantaneous information about the direction of
motion in the vicinity of target
Specifies regions which are more attractive in
nature exit locations
Specifies regions in the space which are more
repulsive in nature. Walls, No-go areas
20Algorithmic Overview
Crowd Video
Probability Computation
21Static Floor Field (SFF)
- Specifies regions of the scene which are more
attractive in nature - exit locations
- Captures static properties of the scene.
- Computed only once.
22Static Floor Field (SFF)
- Attractive Regions Sinks
- If we have the knowledge about the sinks
- For any point in the scene, we can compute
tendency of the individual at that point to move
towards the sink. - Local Force Function of shortest distance to
sink in term of some metric - Distance Metric Number of steps to wade through
the point flow field of the scene.
23Static Floor Field (SFF)
- Two step procedure
- Computation of Point Flows
- Sink Seeking
24Point Flows
- Point flow instantaneous motion at any locations
- location and velocity (Zi(Xi, Vi))
- Motion in a single frame a set of point flows
- Motion in a video point flow field
25Point Flows
- Point flow instantaneous motion at any locations
- location and velocity (Zi(Xi, Vi))
26Sink Seeking
Sink seeking process
Sink path
sink
- Property a particle dropped in the point flow
field will shift to a sink
27Shift Estimation
K1
Vik1
K
t1
n
Given a point i Zi(Xi, Vi), Compute
Zik1 Xik1 Xik Vik Vik1 mean shift on the
neighbors
28video
29Sinks
Sink seeking
30Example SFF
31Example SFF
32Dynamic Floor Field (DFF)
- Instantaneous information about the direction of
motion. - Instantaneous motion constraints likely
locations. - Based on particle advection.
33Dynamic Floor Field (DFF)
- Use particle advection through the instantaneous
flow field. - Count number of particles that pass through
location x and y - Number of common particles Strength of
association between x and y
34Example Dynamic Floor Field
35Boundary Floor Field (BFF)
- Specifies regions in the space which are more
repulsive in nature. - Walls, No-go areas etc.
- Crowd flow segmentation is used to generate it.
36Crowd Flow Segmentation (CVPR2007)
Input Video
37Optical Flow Computation
- Two Schemes
- Block based correlation in Fourier domain
- Locate peaks in the correlation surface for
displacement calculation. - Adaptive local median filtering for outlier
removal. - Thomas Brox et al., High accuracy optical flow
Estimation based on a theory for warping, ECCV
2004 - Grey value constancy Gradient Constancy
Smoothness Multi-scale
38Optical Flow - Examples
Color Coded Optical Flows
Video
39Optical Flow - Examples
Color Coded Optical Flows
Videos
40Particle Flow Map
?
- Keeps track of evolution of particles
- Numerical Integration
- Two-dimensional Phase Space Two Flow Maps
41Particle Flow Map - contd
..
t0
t1
tn
42Particle Flow Maps
Forward X Particle Flow Map
Forward Y Particle Flow Map
Backward X Particle Flow Map
Backward Y Particle Flow Map
43Particle Flow Maps
Forward X Particle Flow Map
Forward Y Particle Flow Map
Backward X Particle Flow Map
Backward Y Particle Flow Map
44Spatial Gradients of Flow Maps
- Observation
- High gradient between particles of different
dynamics in the flow maps. - Belong to dynamically distinct crowd flows.
- Spatial gradients will emphasize locations of
such particles.
45Flow Maps Spatial Gradients
46Flow Map Spatial Gradients
47Finite Time Lyapunov Exponent Field
FTLE Field
- LCS appear as ridges in FTLE field.
- FTLE field computed using the spatial gradients
of the flow maps.
48Lyapunov Exponent
- Lyapunov Exponent
- Asymptotic quantity.
- Measures divergence of nearby particles.
- Measures chaoticity of the system.
- For a given dynamical system
49Finite Time Lyapunov Exponent Field
- Analysis over a grid of particles
?
Magnitude of the separation between the particles
Seek to maximize over all possible choices of
50Finite Time Lyapunov Exponent Field
Using the operator norm
Finite version of Cauchy-Green deformation tensor
51Finite Time Lyapunov Exponent Field
FTLE Field
- Coherent motion within a flow segment.
- Eigen-values close to zero.
- Incoherent motion at the boundary.
- Higher Eigen-values.
- Higher values - Ridges in the FTLE field -
Ridges point to locations of LCS.
52FTLE Field Example 1
53FTLE Field Example 2
54FTLE Field Example 3
55FTLE Field Example 3
FTLE Field Example 4
56FTLE Field Example 2
FTLE Field Example 5
57FTLE Field Segmentation
FTLE Field
FTLE Field
FTLE Field
- Ridges in FTLE field - Watershed lines dividing
individual catchment basins. - Region growing
- Watershed Segmentation Algorithm
58Segmentation Result-1
National Geographic Documentary - Inside Mecca
Using both Forward and Backward FTLE
59Segmentation Result-2
Segmentation Result-2
National Geographic Documentary - Inside Mecca
Video
Using both Forward and Backward FTLE
60Segmentation Result-2
61Boundary Floor Field
62Example Boundary Floor Fields
63Tracking
64Experiment - 1
- Average chip size 14 x 22 pixels
- 492 Frames
- Selected 199 athletes for tracking
- Successfully tracked 143 athletes
65Experiment - 1
66Experiment -1
67Experiment 1
68Qualitative Results
69Qualitative Results
70Experiment-2
- Average chip size 13 x 16 pixels
- 333 Frames
- Selected 120 athletes for tracking
- Successfully tracked 117 athletes
71Experiment - 2
72Experiment -2
73Experiment 2
74Qualitative Results
75Experiment-3
- Average chip size 14 x 17 pixels
- 453 Frames
- Selected 50 athletes for tracking
76Experiment - 3
77Experiment 3
78Qualitative Results
79All Tracks
Successes
80Failures
- Reasons
- Severe occlusion
- Illumination Changes
81Quantitative Analysis
- Ground truth generated for
- 50 athletes from seq-1.
- 20 athletes from seq-2.
- 15 athletes from seq-3.
82Quantitative Analysis
- Ground truth generated for
- 50 athletes from seq-1.
- 20 athletes from seq-2.
- 15 athletes from seq-3.
83Quantitative Analysis
- Ground truth generated for
- 50 athletes from seq-1.
- 20 athletes from seq-2.
- 15 athletes from seq-3.
84Contribution of Fields
- Analysis
- SFF commits less error when motion in straight
path, no side ways - DFF commits more error when athletes overtake
each other
85Summary
Tracking in Crowded Scenes
- Used knowledge of the scene for tracking.
- Concept of floor fields.
- Demonstrated tracking on challenging scenarios.
86Thank YOu
87Tracking Against the Flow
88PART III Motion and Appearance Contexts for
Re-Acquiring Targets
89PART III Motion and Appearance Contexts for
Re-Acquiring Targets