Goal: Fast and Robust Velocity Es5ma5on
P1 P2
P3 P4
Our Approach: Alignment Probability
● Spatial Distance ● Color Distance (if available) ● Probability of Occlusion
Annealed Dynamic Histograms
Combining 3D Shape, Color, and Motion for Robust Anytime Tracking
David Held, Jesse Levinson, Sebas5an Thrun, and Silvio Savarese
p(xt | z1 . . . zt) = η p(zt | xt, zt−1)p(xt | z1 . . . zt−1)
Goal: Fast and Robust Velocity Es5ma5on
Combining 3D Shape, Color, and Motion for Robust Anytime Tracking
David Held, Jesse Levinson, Sebas5an Thrun, and Silvio Savarese
p(xt | z1 . . . zt) = η p(zt | xt, zt−1)p(xt | z1 . . . zt−1)Baseline: Centroid Kalman Filter
Local Search Poor Local Optimum!
t+1 t
Baseline: ICP
Annealed Dynamic Histograms
Goal: Fast and Robust Velocity Es5ma5on
P1 P2
P3 P4
Our Approach: Alignment Probability
● Spatial Distance ● Color Distance (if available) ● Probability of Occlusion
Combining 3D Shape, Color, and Motion for Robust Anytime Tracking
David Held, Jesse Levinson, Sebas5an Thrun, and Silvio Savarese
p(xt | z1 . . . zt) = η p(zt | xt, zt−1)p(xt | z1 . . . zt−1)
Annealed Dynamic Histograms
Mo#va#on Quickly and robustly estimate the speed of nearby objects
Laser Data
Camera Images System
Laser Data
Camera Images System
Previous Work (Teichman, et al)
System
Laser Data
Camera Images
This Work
Velocity Estimation
Previous Work (Teichman, et al)
Velocity Es#ma#on
t
Velocity Es#ma#on
t+1 t
Velocity Es#ma#on
t+1 t
Velocity Es#ma#on
t+1 t
Velocity Es#ma#on
t+1 t
ICP Baseline
ICP Baseline
ICP Baseline
ICP Baseline
Local Search Poor Local Optimum!
ICP Baseline
Tracking Probability
Velocity Es#ma#on
t
Velocity Es#ma#on
t+1 t
Velocity Es#ma#on
t+1 t
Velocity Es#ma#on
t+1 t
Velocity Es#ma#on
t+1 t
Velocity Es#ma#on
t+1 t
Velocity Es#ma#on
t+1 t
Xt
Velocity Es#ma#on
t+1 t
Xt
Measurement Model
Motion Model
Tracking Probability
Measurement Model
Motion Model
Tracking Probability
Constant velocity Kalman filter
Measurement Model
Tracking Probability
Motion Model
Measurement Model
Tracking Probability
Motion Model
Measurement Model
Tracking Probability
Motion Model
Measurement Model
Tracking Probability
Motion Model
Measurement Model
Tracking Probability
Motion Model
Measurement Model
Tracking Probability
Motion Model
Measurement Model
Tracking Probability
Motion Model
Measurement Model
Tracking Probability
Motion Model
Measurement Model
Tracking Probability
Motion Model
Measurement Model
Tracking Probability
Motion Model
Measurement Model
Tracking Probability
Motion Model
Measurement Model
Tracking Probability
Motion Model
k
Measurement Model
Tracking Probability
Motion Model
Sensor noise
Sensor resolution
k
Delta Color Value
Pro
babi
lity
Color Probability
Including Color
Including Color
η!
"
zi∈ztexp(−1
2(zi − z̄j)TΣ−1(zi − z̄j)) + k
#
Including Color
η!
"
zi∈ztexp(−1
2(zi − z̄j)TΣ−1(zi − z̄j)) + k
#
Including Color
Including Color
Including Color
Delta Color Value
Pro
babi
lity
Including Color
Delta Color Value
Pro
babi
lity
Including Color
Delta Color Value
Pro
babi
lity
Probabilis#c Framework
3D Shape Color
Tracking
Motion History
Tracking Probability P1 P2
P3 P4
Tracking Probability vy
vx
??
? ?
?
Tracking Probability vy
vx
Dynamic Decomposi#on vy
vx
Dynamic Decomposi#on vy
vx
Dynamic Decomposi#on vy
vx
Dynamic Decomposi#on vy
vx
Derived from minimizing KL-divergence between approximate distribution and true posterior
Annealing
Inflate the measurement model
Annealing
Inflate the measurement model
Annealing
Inflate the measurement model
Algorithm 1. For each hypothesis
A. Compute the probability of the alignment
Measurement Model
Motion Model
Algorithm 1. For each hypothesis
A. Compute the probability of the alignment
B. Finely sample high probability regions
Measurement Model
Motion Model
Algorithm 1. For each hypothesis
A. Compute the probability of the alignment
B. Finely sample high probability regions
C. Go to step 1 to compute the probability of new hypotheses
Measurement Model
Motion Model
Annealing
More time More accurate
Any#me Tracker
0 0.5 1 1.5 2 2.50
0.1
0.2
0.3
0.4
0.5
0.6
Mean runtime (ms)
RM
S er
ror (
m/s
)
Any#me Tracker
Choose runtime based on: Total runtime requirements Importance of tracked object ...
0 0.5 1 1.5 2 2.50
0.1
0.2
0.3
0.4
0.5
0.6
Mean runtime (ms)
RM
S er
ror (
m/s
)
Comparisons
0 0.5 1 1.5 2 2.50
0.1
0.2
0.3
0.4
0.5
0.6
Mean runtime (ms)
RM
S er
ror (
m/s
)
0 0.5 1 1.5 2 2.50
0.1
0.2
0.3
0.4
0.5
0.6
Mean runtime (ms)
RMS
erro
r (m
/s)
0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2R
MS
erro
r (m
/s)
Mean runtime (ms)0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2R
MS
erro
r (m
/s)
Mean runtime (ms)
Kalman FilterICPKalman ICP with Centroid InitKalman ICP with Kalman InitAnnealed Dynamic Histograms
0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2
RM
S er
ror (
m/s
)
Mean runtime (ms)
Kalman FilterICPKalman ICP with Centroid InitKalman ICP with Kalman InitAnnealed Dynamic Histograms
0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2R
MS
erro
r (m
/s)
Mean runtime (ms)
Comparisons
0 0.5 1 1.5 2 2.50
0.1
0.2
0.3
0.4
0.5
0.6
Mean runtime (ms)
RM
S er
ror (
m/s
)
0 0.5 1 1.5 2 2.50
0.1
0.2
0.3
0.4
0.5
0.6
Mean runtime (ms)
RMS
erro
r (m
/s)
0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2R
MS
erro
r (m
/s)
Mean runtime (ms)0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2R
MS
erro
r (m
/s)
Mean runtime (ms)
Kalman FilterICPKalman ICP with Centroid InitKalman ICP with Kalman InitAnnealed Dynamic Histograms
0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2
RM
S er
ror (
m/s
)
Mean runtime (ms)
0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2
RM
S er
ror (
m/s
)
Mean runtime (ms)
Kalman FilterICPKalman ICP with Centroid InitKalman ICP with Kalman InitAnnealed Dynamic Histograms
0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2R
MS
erro
r (m
/s)
Mean runtime (ms)
Comparisons
0 0.5 1 1.5 2 2.50
0.1
0.2
0.3
0.4
0.5
0.6
Mean runtime (ms)
RM
S er
ror (
m/s
)
0 0.5 1 1.5 2 2.50
0.1
0.2
0.3
0.4
0.5
0.6
Mean runtime (ms)
RMS
erro
r (m
/s)
0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2R
MS
erro
r (m
/s)
Mean runtime (ms)0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2R
MS
erro
r (m
/s)
Mean runtime (ms)0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2R
MS
erro
r (m
/s)
Mean runtime (ms)0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2R
MS
erro
r (m
/s)
Mean runtime (ms)0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2R
MS
erro
r (m
/s)
Mean runtime (ms)
Kalman FilterICPKalman ICP with Centroid InitKalman ICP with Kalman InitAnnealed Dynamic Histograms
0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2
RM
S er
ror (
m/s
)
Mean runtime (ms)
0 0.5 1 1.5 2 2.50
0.2
0.4
0.6
0.8
1
1.2R
MS
erro
r (m
/s)
Mean runtime (ms)
Kalman Filter
Kalman Filter
ADH Tracker (Ours)
Models
Quan#ta#ve Evalua#on 2
People Bikes Cars0
0.1
0.2
0.3
0.4
Crispness
People Bikes Cars0
5
10
Crisp
ness
Kalman FilterKalman ICPADH Tracker (Ours)
Sampling Strategies
0 5 10 15 20 250
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8R
MS
erro
r (m
/s)
Mean runtime (ms)
ADH Tracker (Ours)Dense samplingDense sampling with motion predictionTop cell sampling
Advantages over Radar
Conclusions 3D Shape Color
Tracking
Motion History
● Robust to Occlusions, Viewpoint Changes
Conclusions 3D Shape Color
Tracking
Motion History
● Robust to Occlusions, Viewpoint Changes
● Runs in Real-time ● Robust to Initialization Errors
Delta Color Value
Pro
babi
lity
Color Probability
Error vs Number of Points
0 200 400 600 8000
0.5
1
1.5
Number of points
RMS
erro
r (m
/s)
Kalman FilterADH Tracker (Ours)
Error vs Distance
0 20 40 6000.20.40.60.8
11.21.41.61.8
2
Distance to tracked vehicle (m)
RMS
erro
r (m
/s)
Kalman FilterADH Tracker (Ours)
Error vs Number of Frames
0 100 200 3000
0.5
1
1.5
2
Number of frames tracked
RM
S er
ror (
m/s
)
Error vs Number of Frames
0 5 10 15 20 25 300
0.5
1
1.5
2
Number of frames tracked
RM
S er
ror (
m/s
)
ADH TrackerADH Tracker without motion model
No color No motion model No 3D shape0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
RM
S er
ror (
m/s
)