Vehicle Trajectory Prediction in Crowded Highway Scenarios Using - - PowerPoint PPT Presentation

▶
vehicle trajectory prediction in crowded highway
SMART_READER_LITE
LIVE PREVIEW

Vehicle Trajectory Prediction in Crowded Highway Scenarios Using - - PowerPoint PPT Presentation

May 21, 2023 137 likes •1.28k views

IEEE Intelligent Transportation Systems Conference - ITSC 2020 Vehicle Trajectory Prediction in Crowded Highway Scenarios Using Bird Eye View Representations and CNNs R. Izquierdo, A. Quintanar I. Parra, D. Fernndez-Llorca and M. A. Sotelo

slide-1
SLIDE 1

IEEE Intelligent Transportation Systems Conference - ITSC 2020

Vehicle Trajectory Prediction in Crowded Highway Scenarios Using Bird Eye View Representations and CNNs

  • R. Izquierdo, A. Quintanar
  • I. Parra, D. Fernández-Llorca and M. A. Sotelo

Computer Engineering Department Universidad de Alcalá

20th September 2020

slide-2
SLIDE 2

1

Outline

Motivation Dataset Network Architecture Results Conclusions & Future Work

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-3
SLIDE 3

2

Outline

Motivation Dataset Network Architecture Results Conclusions & Future Work

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-4
SLIDE 4

3

Motivation

Vehicle Trajectory Prediction Problem

  • Structured or unstructured scenarios
  • Multi-agent problem
  • What variables should be used?
  • How can it be modeled?
  • Kinematic or dinamic models
  • Rigid interaction-aware models
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-5
SLIDE 5

3

Motivation

Vehicle Trajectory Prediction Problem

  • Structured or unstructured scenarios
  • Multi-agent problem
  • What variables should be used?
  • How can it be modeled?
  • Kinematic or dinamic models
  • Rigid interaction-aware models
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-6
SLIDE 6

3

Motivation

Vehicle Trajectory Prediction Problem

  • Structured or unstructured scenarios
  • Multi-agent problem
  • What variables should be used?
  • How can it be modeled?
  • Kinematic or dinamic models
  • Rigid interaction-aware models
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-7
SLIDE 7

3

Motivation

Vehicle Trajectory Prediction Problem

  • Structured or unstructured scenarios
  • Multi-agent problem
  • What variables should be used?
  • How can it be modeled?
  • Kinematic or dinamic models
  • Rigid interaction-aware models
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-8
SLIDE 8

3

Motivation

Vehicle Trajectory Prediction Problem

  • Structured or unstructured scenarios
  • Multi-agent problem
  • What variables should be used?
  • How can it be modeled?
  • Kinematic or dinamic models
  • Rigid interaction-aware models
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-9
SLIDE 9

3

Motivation

Vehicle Trajectory Prediction Problem

  • Structured or unstructured scenarios
  • Multi-agent problem
  • What variables should be used?
  • How can it be modeled?
  • Kinematic or dinamic models
  • Rigid interaction-aware models
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-10
SLIDE 10

3

Motivation

Vehicle Trajectory Prediction Problem

  • Structured or unstructured scenarios
  • Multi-agent problem
  • What variables should be used?
  • How can it be modeled?
  • Kinematic or dinamic models
  • Rigid interaction-aware models
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-11
SLIDE 11

4

Motivation

Our proposal

A deep learning-based trajectory prediction approach based on graphic representations

  • Motion and interaction histories
  • Allows context integration
  • Unlimited in number of vehicles
  • No vehicle-centered
  • Simultaneous prediction
  • Unlimited but fixed range and prediction horizon
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-12
SLIDE 12

4

Motivation

Our proposal

A deep learning-based trajectory prediction approach based on graphic representations

  • Motion and interaction histories
  • Allows context integration
  • Unlimited in number of vehicles
  • No vehicle-centered
  • Simultaneous prediction
  • Unlimited but fixed range and prediction horizon
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-13
SLIDE 13

4

Motivation

Our proposal

A deep learning-based trajectory prediction approach based on graphic representations

  • Motion and interaction histories
  • Allows context integration
  • Unlimited in number of vehicles
  • No vehicle-centered
  • Simultaneous prediction
  • Unlimited but fixed range and prediction horizon
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-14
SLIDE 14

4

Motivation

Our proposal

A deep learning-based trajectory prediction approach based on graphic representations

  • Motion and interaction histories
  • Allows context integration
  • Unlimited in number of vehicles
  • No vehicle-centered
  • Simultaneous prediction
  • Unlimited but fixed range and prediction horizon
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-15
SLIDE 15

4

Motivation

Our proposal

A deep learning-based trajectory prediction approach based on graphic representations

  • Motion and interaction histories
  • Allows context integration
  • Unlimited in number of vehicles
  • No vehicle-centered
  • Simultaneous prediction
  • Unlimited but fixed range and prediction horizon
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-16
SLIDE 16

4

Motivation

Our proposal

A deep learning-based trajectory prediction approach based on graphic representations

  • Motion and interaction histories
  • Allows context integration
  • Unlimited in number of vehicles
  • No vehicle-centered
  • Simultaneous prediction
  • Unlimited but fixed range and prediction horizon
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-17
SLIDE 17

4

Motivation

Our proposal

A deep learning-based trajectory prediction approach based on graphic representations

  • Motion and interaction histories
  • Allows context integration
  • Unlimited in number of vehicles
  • No vehicle-centered
  • Simultaneous prediction
  • Unlimited but fixed range and prediction horizon
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-18
SLIDE 18

4

Motivation

Our proposal

A deep learning-based trajectory prediction approach based on graphic representations

  • Motion and interaction histories
  • Allows context integration
  • Unlimited in number of vehicles
  • No vehicle-centered
  • Simultaneous prediction
  • Unlimited but fixed range and prediction horizon
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-19
SLIDE 19

5

Outline

Motivation Dataset Network Architecture Results Conclusions & Future Work

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-20
SLIDE 20

6

Dataset

HighD Dataset1

  • Publicy available at highd-dataset.com
  • Road Structure
  • Unique IDs
  • Vehicle dimension
  • Position, Speed, and Acceleration at 25 Hz

HighD Challenge Samples 39M 1500K Vehicles 110K 110K Hours 147 16 Clips 60 60

1The highD Dataset: A Drone Dataset of Naturalistic Vehicle Trajectories on German Highways for Validation of Highly Automated Driving Systems. Krajewski, Robert and

Bock, Julian and Kloeker, Laurent and Eckstein, Lutz

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-21
SLIDE 21

6

Dataset

HighD Dataset1

  • Publicy available at highd-dataset.com
  • Road Structure
  • Unique IDs
  • Vehicle dimension
  • Position, Speed, and Acceleration at 25 Hz

HighD Challenge Samples 39M 1500K Vehicles 110K 110K Hours 147 16 Clips 60 60

1The highD Dataset: A Drone Dataset of Naturalistic Vehicle Trajectories on German Highways for Validation of Highly Automated Driving Systems. Krajewski, Robert and

Bock, Julian and Kloeker, Laurent and Eckstein, Lutz

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-22
SLIDE 22

6

Dataset

HighD Dataset1

  • Publicy available at highd-dataset.com
  • Road Structure
  • Unique IDs
  • Vehicle dimension
  • Position, Speed, and Acceleration at 25 Hz

HighD Challenge Samples 39M 1500K Vehicles 110K 110K Hours 147 16 Clips 60 60

1The highD Dataset: A Drone Dataset of Naturalistic Vehicle Trajectories on German Highways for Validation of Highly Automated Driving Systems. Krajewski, Robert and

Bock, Julian and Kloeker, Laurent and Eckstein, Lutz

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-23
SLIDE 23

6

Dataset

HighD Dataset1

  • Publicy available at highd-dataset.com
  • Road Structure
  • Unique IDs
  • Vehicle dimension
  • Position, Speed, and Acceleration at 25 Hz

HighD Challenge Samples 39M 1500K Vehicles 110K 110K Hours 147 16 Clips 60 60

1The highD Dataset: A Drone Dataset of Naturalistic Vehicle Trajectories on German Highways for Validation of Highly Automated Driving Systems. Krajewski, Robert and

Bock, Julian and Kloeker, Laurent and Eckstein, Lutz

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-24
SLIDE 24

6

Dataset

HighD Dataset1

  • Publicy available at highd-dataset.com
  • Road Structure
  • Unique IDs
  • Vehicle dimension
  • Position, Speed, and Acceleration at 25 Hz

HighD Challenge Samples 39M 1500K Vehicles 110K 110K Hours 147 16 Clips 60 60

1The highD Dataset: A Drone Dataset of Naturalistic Vehicle Trajectories on German Highways for Validation of Highly Automated Driving Systems. Krajewski, Robert and

Bock, Julian and Kloeker, Laurent and Eckstein, Lutz

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-25
SLIDE 25

6

Dataset

HighD Dataset1

  • Publicy available at highd-dataset.com
  • Road Structure
  • Unique IDs
  • Vehicle dimension
  • Position, Speed, and Acceleration at 25 Hz

HighD Challenge Samples 39M 1500K Vehicles 110K 110K Hours 147 16 Clips 60 60

1The highD Dataset: A Drone Dataset of Naturalistic Vehicle Trajectories on German Highways for Validation of Highly Automated Driving Systems. Krajewski, Robert and

Bock, Julian and Kloeker, Laurent and Eckstein, Lutz

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-26
SLIDE 26

6

Dataset

HighD Dataset1

  • Publicy available at highd-dataset.com
  • Road Structure
  • Unique IDs
  • Vehicle dimension
  • Position, Speed, and Acceleration at 25 Hz

HighD Challenge Samples 39M 1500K Vehicles 110K 110K Hours 147 16 Clips 60 60

1The highD Dataset: A Drone Dataset of Naturalistic Vehicle Trajectories on German Highways for Validation of Highly Automated Driving Systems. Krajewski, Robert and

Bock, Julian and Kloeker, Laurent and Eckstein, Lutz

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-27
SLIDE 27

6

Dataset

HighD Dataset1

  • Publicy available at highd-dataset.com
  • Road Structure
  • Unique IDs
  • Vehicle dimension
  • Position, Speed, and Acceleration at 25 Hz

HighD Challenge Samples 39M 1500K Vehicles 110K 110K Hours 147 16 Clips 60 60

1The highD Dataset: A Drone Dataset of Naturalistic Vehicle Trajectories on German Highways for Validation of Highly Automated Driving Systems. Krajewski, Robert and

Bock, Julian and Kloeker, Laurent and Eckstein, Lutz

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-28
SLIDE 28

6

Dataset

HighD Dataset1

  • Publicy available at highd-dataset.com
  • Road Structure
  • Unique IDs
  • Vehicle dimension
  • Position, Speed, and Acceleration at 25 Hz

HighD Challenge Samples 39M 1500K Vehicles 110K 110K Hours 147 16 Clips 60 60

1The highD Dataset: A Drone Dataset of Naturalistic Vehicle Trajectories on German Highways for Validation of Highly Automated Driving Systems. Krajewski, Robert and

Bock, Julian and Kloeker, Laurent and Eckstein, Lutz

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-29
SLIDE 29

6

Dataset

HighD Dataset1

  • Publicy available at highd-dataset.com
  • Road Structure
  • Unique IDs
  • Vehicle dimension
  • Position, Speed, and Acceleration at 25 Hz

HighD Challenge Samples 39M 1500K Vehicles 110K 110K Hours 147 16 Clips 60 60

1The highD Dataset: A Drone Dataset of Naturalistic Vehicle Trajectories on German Highways for Validation of Highly Automated Driving Systems. Krajewski, Robert and

Bock, Julian and Kloeker, Laurent and Eckstein, Lutz

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-30
SLIDE 30

7

Dataset

Data Codification

  • Driving Area as an Image: 32x512 m. → 64x512 px.
  • Vehicles as N(µ, σ2)
  • Possible Overlap P(x, y) = max {Ni} ∀i
  • Only Predictable Vehicles
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-31
SLIDE 31

7

Dataset

Data Codification

  • Driving Area as an Image: 32x512 m. → 64x512 px.
  • Vehicles as N(µ, σ2)
  • Possible Overlap P(x, y) = max {Ni} ∀i
  • Only Predictable Vehicles
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-32
SLIDE 32

7

Dataset

Data Codification

  • Driving Area as an Image: 32x512 m. → 64x512 px.
  • Vehicles as N(µ, σ2)
  • Possible Overlap P(x, y) = max {Ni} ∀i
  • Only Predictable Vehicles
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-33
SLIDE 33

7

Dataset

Data Codification

  • Driving Area as an Image: 32x512 m. → 64x512 px.
  • Vehicles as N(µ, σ2)
  • Possible Overlap P(x, y) = max {Ni} ∀i
  • Only Predictable Vehicles
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-34
SLIDE 34

7

Dataset

Data Codification

  • Driving Area as an Image: 32x512 m. → 64x512 px.
  • Vehicles as N(µ, σ2)
  • Possible Overlap P(x, y) = max {Ni} ∀i
  • Only Predictable Vehicles
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-35
SLIDE 35

8

Dataset

Data Decodification

  • Iterative Extraction of Vehicle Positions
  • Search the pixel with the higest probability
  • Compute the mass center
  • Reset the area
  • Vehicle association by Euclidean distance
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-36
SLIDE 36

8

Dataset

Data Decodification

  • Iterative Extraction of Vehicle Positions
  • Search the pixel with the higest probability
  • Compute the mass center
  • Reset the area
  • Vehicle association by Euclidean distance
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-37
SLIDE 37

8

Dataset

Data Decodification

  • Iterative Extraction of Vehicle Positions
  • Search the pixel with the higest probability
  • Compute the mass center
  • Reset the area
  • Vehicle association by Euclidean distance
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-38
SLIDE 38

8

Dataset

Data Decodification

  • Iterative Extraction of Vehicle Positions
  • Search the pixel with the higest probability
  • Compute the mass center
  • Reset the area
  • Vehicle association by Euclidean distance
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-39
SLIDE 39

8

Dataset

Data Decodification

  • Iterative Extraction of Vehicle Positions
  • Search the pixel with the higest probability
  • Compute the mass center
  • Reset the area
  • Vehicle association by Euclidean distance
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-40
SLIDE 40

8

Dataset

Data Decodification

  • Iterative Extraction of Vehicle Positions
  • Search the pixel with the higest probability
  • Compute the mass center
  • Reset the area
  • Vehicle association by Euclidean distance
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-41
SLIDE 41

8

Dataset

Data Decodification

  • Iterative Extraction of Vehicle Positions
  • Search the pixel with the higest probability
  • Compute the mass center
  • Reset the area
  • Vehicle association by Euclidean distance
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-42
SLIDE 42

8

Dataset

Data Decodification

  • Iterative Extraction of Vehicle Positions
  • Search the pixel with the higest probability
  • Compute the mass center
  • Reset the area
  • Vehicle association by Euclidean distance
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-43
SLIDE 43

8

Dataset

Data Decodification

  • Iterative Extraction of Vehicle Positions
  • Search the pixel with the higest probability
  • Compute the mass center
  • Reset the area
  • Vehicle association by Euclidean distance
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-44
SLIDE 44

8

Dataset

Data Decodification

  • Iterative Extraction of Vehicle Positions
  • Search the pixel with the higest probability
  • Compute the mass center
  • Reset the area
  • Vehicle association by Euclidean distance
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-45
SLIDE 45

9

Outline

Motivation Dataset Network Architecture Results Conclusions & Future Work

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-46
SLIDE 46

10

Network Architecture

U-net model

  • Number of filters.
  • Depth levels.

Depth levels 4 5 6 7 Receptive Field ±76 ±156 ±316 ±636 Input size 16 32 64 128 Parameters 56k 116k 235k 472k

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-47
SLIDE 47

10

Network Architecture

U-net model

  • Number of filters.
  • Depth levels.

Depth levels 4 5 6 7 Receptive Field ±76 ±156 ±316 ±636 Input size 16 32 64 128 Parameters 56k 116k 235k 472k

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-48
SLIDE 48

10

Network Architecture

U-net model

  • Number of filters.
  • Depth levels.

Depth levels 4 5 6 7 Receptive Field ±76 ±156 ±316 ±636 Input size 16 32 64 128 Parameters 56k 116k 235k 472k

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-49
SLIDE 49

10

Network Architecture

U-net model

  • Number of filters.
  • Depth levels.

Depth levels 4 5 6 7 Receptive Field ±76 ±156 ±316 ±636 Input size 16 32 64 128 Parameters 56k 116k 235k 472k

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-50
SLIDE 50

10

Network Architecture

U-net model

  • Number of filters.
  • Depth levels.

Depth levels 4 5 6 7 Receptive Field ±76 ±156 ±316 ±636 Input size 16 32 64 128 Parameters 56k 116k 235k 472k

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-51
SLIDE 51

10

Network Architecture

U-net model

  • Number of filters.
  • Depth levels.

Depth levels 4 5 6 7 Receptive Field ±76 ±156 ±316 ±636 Input size 16 32 64 128 Parameters 56k 116k 235k 472k

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-52
SLIDE 52

10

Network Architecture

U-net model

  • Number of filters.
  • Depth levels.

Depth levels 4 5 6 7 Receptive Field ±76 ±156 ±316 ±636 Input size 16 32 64 128 Parameters 56k 116k 235k 472k

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-53
SLIDE 53

11

Network Architecture

Problem Application

  • Data rate lowered to 5 Hz
  • Input & Output

15@64x512 BEV −2.8 ≤ t ≤ 3.0

  • Activation Layer:
  • Linear
  • Clipped Rect. Linear Unit
  • Hyperbolic tangent
  • Loss function SE
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-54
SLIDE 54

11

Network Architecture

Problem Application

  • Data rate lowered to 5 Hz
  • Input & Output

15@64x512 BEV −2.8 ≤ t ≤ 3.0

  • Activation Layer:
  • Linear
  • Clipped Rect. Linear Unit
  • Hyperbolic tangent
  • Loss function SE
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-55
SLIDE 55

11

Network Architecture

Problem Application

  • Data rate lowered to 5 Hz
  • Input & Output

15@64x512 BEV −2.8 ≤ t ≤ 3.0

  • Activation Layer:
  • Linear
  • Clipped Rect. Linear Unit
  • Hyperbolic tangent
  • Loss function SE
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-56
SLIDE 56

11

Network Architecture

Problem Application

  • Data rate lowered to 5 Hz
  • Input & Output

15@64x512 BEV −2.8 ≤ t ≤ 3.0

  • Activation Layer:
  • Linear
  • Clipped Rect. Linear Unit
  • Hyperbolic tangent
  • Loss function SE
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-57
SLIDE 57

11

Network Architecture

Problem Application

  • Data rate lowered to 5 Hz
  • Input & Output

15@64x512 BEV −2.8 ≤ t ≤ 3.0

  • Activation Layer:
  • Linear
  • Clipped Rect. Linear Unit
  • Hyperbolic tangent
  • Loss function SE
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-58
SLIDE 58

11

Network Architecture

Problem Application

  • Data rate lowered to 5 Hz
  • Input & Output

15@64x512 BEV −2.8 ≤ t ≤ 3.0

  • Activation Layer:
  • Linear
  • Clipped Rect. Linear Unit
  • Hyperbolic tangent
  • Loss function SE
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-59
SLIDE 59

11

Network Architecture

Problem Application

  • Data rate lowered to 5 Hz
  • Input & Output

15@64x512 BEV −2.8 ≤ t ≤ 3.0

  • Activation Layer:
  • Linear
  • Clipped Rect. Linear Unit
  • Hyperbolic tangent
  • Loss function SE
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-60
SLIDE 60

11

Network Architecture

Problem Application

  • Data rate lowered to 5 Hz
  • Input & Output

15@64x512 BEV −2.8 ≤ t ≤ 3.0

  • Activation Layer:
  • Linear
  • Clipped Rect. Linear Unit
  • Hyperbolic tangent
  • Loss function SE
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-61
SLIDE 61

11

Network Architecture

Problem Application

  • Data rate lowered to 5 Hz
  • Input & Output

15@64x512 BEV −2.8 ≤ t ≤ 3.0

  • Activation Layer:
  • Linear
  • Clipped Rect. Linear Unit
  • Hyperbolic tangent
  • Loss function SE
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-62
SLIDE 62

12

Outline

Motivation Dataset Network Architecture Results Conclusions & Future Work

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-63
SLIDE 63

13

Results

Training setup

  • Seq 1-20 28K samples
  • Mini-batch size 1
  • Epoch 1
  • Learning rate 10−6
  • Momentum 0.9
  • Loss function SE

Test Results

  • Seq 21-25 7K samples
  • Position MAE

t = 0.2 t = 3.0 Model εx / εy εx / εy

  • Const. acc.

0.02 / 0.00 0.23 / 0.17 d = 5, f = linear 0.52 / 0.17 2.36 / 0.54 d = 6, f = linear 0.23 / 0.01 1.23 / 0.07 d = 5, f = tanh

  • / -
  • / -

d = 6, f = tanh

  • / -
  • / -

d = 5, f = cRelu 0.74 / 0.38 2.51 / 0.94 d = 6, f = cRelu 0.46 / 0.22 2.06 / 0.62

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-64
SLIDE 64

13

Results

Training setup

  • Seq 1-20 28K samples
  • Mini-batch size 1
  • Epoch 1
  • Learning rate 10−6
  • Momentum 0.9
  • Loss function SE

Test Results

  • Seq 21-25 7K samples
  • Position MAE

t = 0.2 t = 3.0 Model εx / εy εx / εy

  • Const. acc.

0.02 / 0.00 0.23 / 0.17 d = 5, f = linear 0.52 / 0.17 2.36 / 0.54 d = 6, f = linear 0.23 / 0.01 1.23 / 0.07 d = 5, f = tanh

  • / -
  • / -

d = 6, f = tanh

  • / -
  • / -

d = 5, f = cRelu 0.74 / 0.38 2.51 / 0.94 d = 6, f = cRelu 0.46 / 0.22 2.06 / 0.62

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-65
SLIDE 65

13

Results

Training setup

  • Seq 1-20 28K samples
  • Mini-batch size 1
  • Epoch 1
  • Learning rate 10−6
  • Momentum 0.9
  • Loss function SE

Test Results

  • Seq 21-25 7K samples
  • Position MAE

t = 0.2 t = 3.0 Model εx / εy εx / εy

  • Const. acc.

0.02 / 0.00 0.23 / 0.17 d = 5, f = linear 0.52 / 0.17 2.36 / 0.54 d = 6, f = linear 0.23 / 0.01 1.23 / 0.07 d = 5, f = tanh

  • / -
  • / -

d = 6, f = tanh

  • / -
  • / -

d = 5, f = cRelu 0.74 / 0.38 2.51 / 0.94 d = 6, f = cRelu 0.46 / 0.22 2.06 / 0.62

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-66
SLIDE 66

13

Results

Training setup

  • Seq 1-20 28K samples
  • Mini-batch size 1
  • Epoch 1
  • Learning rate 10−6
  • Momentum 0.9
  • Loss function SE

Test Results

  • Seq 21-25 7K samples
  • Position MAE

t = 0.2 t = 3.0 Model εx / εy εx / εy

  • Const. acc.

0.02 / 0.00 0.23 / 0.17 d = 5, f = linear 0.52 / 0.17 2.36 / 0.54 d = 6, f = linear 0.23 / 0.01 1.23 / 0.07 d = 5, f = tanh

  • / -
  • / -

d = 6, f = tanh

  • / -
  • / -

d = 5, f = cRelu 0.74 / 0.38 2.51 / 0.94 d = 6, f = cRelu 0.46 / 0.22 2.06 / 0.62

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-67
SLIDE 67

13

Results

Training setup

  • Seq 1-20 28K samples
  • Mini-batch size 1
  • Epoch 1
  • Learning rate 10−6
  • Momentum 0.9
  • Loss function SE

Test Results

  • Seq 21-25 7K samples
  • Position MAE

t = 0.2 t = 3.0 Model εx / εy εx / εy

  • Const. acc.

0.02 / 0.00 0.23 / 0.17 d = 5, f = linear 0.52 / 0.17 2.36 / 0.54 d = 6, f = linear 0.23 / 0.01 1.23 / 0.07 d = 5, f = tanh

  • / -
  • / -

d = 6, f = tanh

  • / -
  • / -

d = 5, f = cRelu 0.74 / 0.38 2.51 / 0.94 d = 6, f = cRelu 0.46 / 0.22 2.06 / 0.62

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-68
SLIDE 68

13

Results

Training setup

  • Seq 1-20 28K samples
  • Mini-batch size 1
  • Epoch 1
  • Learning rate 10−6
  • Momentum 0.9
  • Loss function SE

Test Results

  • Seq 21-25 7K samples
  • Position MAE

t = 0.2 t = 3.0 Model εx / εy εx / εy

  • Const. acc.

0.02 / 0.00 0.23 / 0.17 d = 5, f = linear 0.52 / 0.17 2.36 / 0.54 d = 6, f = linear 0.23 / 0.01 1.23 / 0.07 d = 5, f = tanh

  • / -
  • / -

d = 6, f = tanh

  • / -
  • / -

d = 5, f = cRelu 0.74 / 0.38 2.51 / 0.94 d = 6, f = cRelu 0.46 / 0.22 2.06 / 0.62

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-69
SLIDE 69

13

Results

Training setup

  • Seq 1-20 28K samples
  • Mini-batch size 1
  • Epoch 1
  • Learning rate 10−6
  • Momentum 0.9
  • Loss function SE

Test Results

  • Seq 21-25 7K samples
  • Position MAE

t = 0.2 t = 3.0 Model εx / εy εx / εy

  • Const. acc.

0.02 / 0.00 0.23 / 0.17 d = 5, f = linear 0.52 / 0.17 2.36 / 0.54 d = 6, f = linear 0.23 / 0.01 1.23 / 0.07 d = 5, f = tanh

  • / -
  • / -

d = 6, f = tanh

  • / -
  • / -

d = 5, f = cRelu 0.74 / 0.38 2.51 / 0.94 d = 6, f = cRelu 0.46 / 0.22 2.06 / 0.62

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-70
SLIDE 70

13

Results

Training setup

  • Seq 1-20 28K samples
  • Mini-batch size 1
  • Epoch 1
  • Learning rate 10−6
  • Momentum 0.9
  • Loss function SE

Test Results

  • Seq 21-25 7K samples
  • Position MAE

t = 0.2 t = 3.0 Model εx / εy εx / εy

  • Const. acc.

0.02 / 0.00 0.23 / 0.17 d = 5, f = linear 0.52 / 0.17 2.36 / 0.54 d = 6, f = linear 0.23 / 0.01 1.23 / 0.07 d = 5, f = tanh

  • / -
  • / -

d = 6, f = tanh

  • / -
  • / -

d = 5, f = cRelu 0.74 / 0.38 2.51 / 0.94 d = 6, f = cRelu 0.46 / 0.22 2.06 / 0.62

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-71
SLIDE 71

13

Results

Training setup

  • Seq 1-20 28K samples
  • Mini-batch size 1
  • Epoch 1
  • Learning rate 10−6
  • Momentum 0.9
  • Loss function SE

Test Results

  • Seq 21-25 7K samples
  • Position MAE

t = 0.2 t = 3.0 Model εx / εy εx / εy

  • Const. acc.

0.02 / 0.00 0.23 / 0.17 d = 5, f = linear 0.52 / 0.17 2.36 / 0.54 d = 6, f = linear 0.23 / 0.01 1.23 / 0.07 d = 5, f = tanh

  • / -
  • / -

d = 6, f = tanh

  • / -
  • / -

d = 5, f = cRelu 0.74 / 0.38 2.51 / 0.94 d = 6, f = cRelu 0.46 / 0.22 2.06 / 0.62

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-72
SLIDE 72

14

Results

Training setup

  • Seq 1-20 28K samples
  • Mini-batch size 1
  • Epoch 1
  • Learning rate 10−6
  • Momentum 0.9
  • Loss function SE

Test Results

  • Seq 21-25 7K samples
  • Position MAE
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-73
SLIDE 73

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-74
SLIDE 74

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-75
SLIDE 75

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-76
SLIDE 76

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-77
SLIDE 77

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-78
SLIDE 78

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-79
SLIDE 79

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-80
SLIDE 80

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-81
SLIDE 81

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-82
SLIDE 82

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-83
SLIDE 83

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-84
SLIDE 84

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-85
SLIDE 85

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-86
SLIDE 86

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-87
SLIDE 87

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-88
SLIDE 88

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-89
SLIDE 89

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-90
SLIDE 90

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-91
SLIDE 91

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-92
SLIDE 92

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-93
SLIDE 93

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-94
SLIDE 94

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-95
SLIDE 95

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-96
SLIDE 96

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-97
SLIDE 97

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-98
SLIDE 98

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-99
SLIDE 99

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-100
SLIDE 100

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-101
SLIDE 101

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-102
SLIDE 102

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-103
SLIDE 103

15

Results

Results Example

Input block example Output block example (× = GT + = Pred)

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-104
SLIDE 104

16

Outline

Motivation Dataset Network Architecture Results Conclusions & Future Work

  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-105
SLIDE 105

17

Conclusions & Future Work

Conclusions

  • The U-net model has been adapted to predict trajectories
  • Prediction are performed simultaneously in an interactive way
  • U-net overcomes const. acc. model in lateral prediction

Future Work

  • Improve the vehicle extraction method
  • Modify parameters to train deeper U-net configurations
  • Apply this approach to intersection and roundabout scenarios
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-106
SLIDE 106

17

Conclusions & Future Work

Conclusions

  • The U-net model has been adapted to predict trajectories
  • Prediction are performed simultaneously in an interactive way
  • U-net overcomes const. acc. model in lateral prediction

Future Work

  • Improve the vehicle extraction method
  • Modify parameters to train deeper U-net configurations
  • Apply this approach to intersection and roundabout scenarios
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-107
SLIDE 107

17

Conclusions & Future Work

Conclusions

  • The U-net model has been adapted to predict trajectories
  • Prediction are performed simultaneously in an interactive way
  • U-net overcomes const. acc. model in lateral prediction

Future Work

  • Improve the vehicle extraction method
  • Modify parameters to train deeper U-net configurations
  • Apply this approach to intersection and roundabout scenarios
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-108
SLIDE 108

17

Conclusions & Future Work

Conclusions

  • The U-net model has been adapted to predict trajectories
  • Prediction are performed simultaneously in an interactive way
  • U-net overcomes const. acc. model in lateral prediction

Future Work

  • Improve the vehicle extraction method
  • Modify parameters to train deeper U-net configurations
  • Apply this approach to intersection and roundabout scenarios
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-109
SLIDE 109

17

Conclusions & Future Work

Conclusions

  • The U-net model has been adapted to predict trajectories
  • Prediction are performed simultaneously in an interactive way
  • U-net overcomes const. acc. model in lateral prediction

Future Work

  • Improve the vehicle extraction method
  • Modify parameters to train deeper U-net configurations
  • Apply this approach to intersection and roundabout scenarios
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-110
SLIDE 110

17

Conclusions & Future Work

Conclusions

  • The U-net model has been adapted to predict trajectories
  • Prediction are performed simultaneously in an interactive way
  • U-net overcomes const. acc. model in lateral prediction

Future Work

  • Improve the vehicle extraction method
  • Modify parameters to train deeper U-net configurations
  • Apply this approach to intersection and roundabout scenarios
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-111
SLIDE 111

17

Conclusions & Future Work

Conclusions

  • The U-net model has been adapted to predict trajectories
  • Prediction are performed simultaneously in an interactive way
  • U-net overcomes const. acc. model in lateral prediction

Future Work

  • Improve the vehicle extraction method
  • Modify parameters to train deeper U-net configurations
  • Apply this approach to intersection and roundabout scenarios
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-112
SLIDE 112

17

Conclusions & Future Work

Conclusions

  • The U-net model has been adapted to predict trajectories
  • Prediction are performed simultaneously in an interactive way
  • U-net overcomes const. acc. model in lateral prediction

Future Work

  • Improve the vehicle extraction method
  • Modify parameters to train deeper U-net configurations
  • Apply this approach to intersection and roundabout scenarios
  • R. Izquierdo | ITSC 2020, Virtual Conference
slide-113
SLIDE 113

IEEE Intelligent Transportation Systems Conference - ITSC 2020

Vehicle Trajectory Prediction in Crowded Highway Scenarios Using Bird Eye View Representations and CNNs

20th September 2020