July 4 - 6, 2022
2 n d E d i t i o n
BigML, Inc #DutchMLSchool
Smart Safe Stadiums
A nice and safe stadium experience for everyone
Ramon van Ingen


CTIO, Siip group
2
Pablo González


Machine Learning Engineer, BigML
BigML, Inc #DutchMLSchool 3
Agenda
• Background


• Our solution


• Smart Video Review Framework
BigML, Inc #DutchMLSchool 4
BigML, Inc #DutchMLSchool 5
“Our football belongs to everyone”
• Attack plan to counter racism and discrimination


• Applying smart technology


• Aligns with a safer, more secure stadium
environment for players, fans, and employees
Background
BigML, Inc #DutchMLSchool 6
A nice and safe stadium experience for everyone
Our solution
BigML, Inc #DutchMLSchool 7
2 innovative solutions
• Access right to stadium is linked to a real
identity through a fan-app (pre-registration)


• Early signalling or detection of unwanted
behaviour or events using AI/ML
Our solution
BigML, Inc #DutchMLSchool 8
Market potential
• All sport stadiums


• Festivals, concerts, and large events


• Easy integration with existing technology,
systems, and processes in stadiums


• Easy connection to ticket-selling agencies
Our solution
Scalability
BigML, Inc #DutchMLSchool 9
Privacy and Ethics by Design
• Pre-registration: Data is securely stored only on individual’s own
device. Only necessary data is shared with consent to attend a
game


• AI: Use of aggregated metadata, no PII or personal
characteristics. Non-discriminatory model that explains what
prediction is based on.


• Segregation of pre-registration and AI: Human in the loop


• Guidance Ethics by University of Applied Science Windesheim
Our solution
BigML, Inc #DutchMLSchool 10
Smart Video Review Framework
Focus your attention on what’s important
!
x2 x3
!
!
BigML, Inc #DutchMLSchool 11
The framework:
fi
nd, group, label, review
Find interesting events
Flags
Chants
Group events
Experts review clusters 

and assign their own labels
1 2
Find anomalous events
Cluster similar events
Smoke bombs / Flares
Label event groups
3
4
Assign labels to
known events
Models learn new
labels from
expert’s feedback
Smart Video Review Framework
BigML, Inc #DutchMLSchool 12
Avigilon Video Security System
The hardware
Smart Video Review Framework
BigML, Inc #DutchMLSchool 13
The software
Smart Video Review Framework
BigML, Inc #DutchMLSchool 14
The problem: We are trying to
fi
nd black swans
The Black Swan: The Impact of the Highly Improbable
Nassim Nicholas Taleb
Smart Video Review Framework
BigML, Inc #DutchMLSchool 15
The problem: We are trying to
fi
nd black swans
A methodology that combines supervised and
unsupervised learning to capture variance in data
Our approach
• Our objective is not to replace the reviewer,
but to make her work easier


• Only the reviewer is able to confidently
identify real black swans
• We will capture the variance in the data
trying to maximize the probability of
finding a black swan
+3 MORE +10 MORE +6 MORE
+8 MORE
+5 MORE
Reviewer’s feedback
+2 MORE
Feature extraction
Clustering
Anomaly Detection
Supervised Models
Anomalous group. It should be
reviewed first
Smart Video Review Framework
BigML, Inc #DutchMLSchool 16
Flare/Smoke Bombs Analysis
Binary Classification: Smoke/Flare vs not Detection: Flare Detection
Smart Video Review Framework
BigML, Inc #DutchMLSchool 17
Flags/Banners Analysis
Detection: Waving-flag and Banner Clustering and Pattern Matching
Smart Video Review Framework
BigML, Inc #DutchMLSchool 18
Chants Analysis
raw microphones data
1 3 seconds sound blocks
2
2D spectrograms
3
Filter out high frequencies
and obtain chromagram
4
Smart Video Review Framework
BigML, Inc #DutchMLSchool 19
Chants Analysis
raw microphones data
1 3 seconds sound blocks
2
2D spectrograms
3
Filter out high frequencies
and obtain chromagram
4
Chromagram Feature
Extraction
5
Chants Clustering and
Fingerprinting Matching
7
Chant / Noise 

Classification
6
Chant 1
Chant 2
Chant 3
Smart Video Review Framework
BigML, Inc #DutchMLSchool 20
Monkey Chants Analysis
raw microphones data
1 3 seconds sound blocks
2
2D spectrograms
3
Filter out high frequencies
and obtain MEL Spectrogram
4
monkey	chant	audio	blocks	
non	monkey	chant	audio	blocks	
Monkey Chant:
Periodic sound
around 400 Hz
BigML Image
Classification model
Smart Video Review Framework
DutchMLSchool 2022 - Smart Safe Stadiums

DutchMLSchool 2022 - Smart Safe Stadiums

  • 1.
    July 4 -6, 2022 2 n d E d i t i o n
  • 2.
    BigML, Inc #DutchMLSchool SmartSafe Stadiums A nice and safe stadium experience for everyone Ramon van Ingen CTIO, Siip group 2 Pablo González Machine Learning Engineer, BigML
  • 3.
    BigML, Inc #DutchMLSchool3 Agenda • Background • Our solution • Smart Video Review Framework
  • 4.
  • 5.
    BigML, Inc #DutchMLSchool5 “Our football belongs to everyone” • Attack plan to counter racism and discrimination • Applying smart technology • Aligns with a safer, more secure stadium environment for players, fans, and employees Background
  • 6.
    BigML, Inc #DutchMLSchool6 A nice and safe stadium experience for everyone Our solution
  • 7.
    BigML, Inc #DutchMLSchool7 2 innovative solutions • Access right to stadium is linked to a real identity through a fan-app (pre-registration) • Early signalling or detection of unwanted behaviour or events using AI/ML Our solution
  • 8.
    BigML, Inc #DutchMLSchool8 Market potential • All sport stadiums • Festivals, concerts, and large events • Easy integration with existing technology, systems, and processes in stadiums • Easy connection to ticket-selling agencies Our solution Scalability
  • 9.
    BigML, Inc #DutchMLSchool9 Privacy and Ethics by Design • Pre-registration: Data is securely stored only on individual’s own device. Only necessary data is shared with consent to attend a game • AI: Use of aggregated metadata, no PII or personal characteristics. Non-discriminatory model that explains what prediction is based on. • Segregation of pre-registration and AI: Human in the loop • Guidance Ethics by University of Applied Science Windesheim Our solution
  • 10.
    BigML, Inc #DutchMLSchool10 Smart Video Review Framework Focus your attention on what’s important ! x2 x3 ! !
  • 11.
    BigML, Inc #DutchMLSchool11 The framework: fi nd, group, label, review Find interesting events Flags Chants Group events Experts review clusters and assign their own labels 1 2 Find anomalous events Cluster similar events Smoke bombs / Flares Label event groups 3 4 Assign labels to known events Models learn new labels from expert’s feedback Smart Video Review Framework
  • 12.
    BigML, Inc #DutchMLSchool12 Avigilon Video Security System The hardware Smart Video Review Framework
  • 13.
    BigML, Inc #DutchMLSchool13 The software Smart Video Review Framework
  • 14.
    BigML, Inc #DutchMLSchool14 The problem: We are trying to fi nd black swans The Black Swan: The Impact of the Highly Improbable Nassim Nicholas Taleb Smart Video Review Framework
  • 15.
    BigML, Inc #DutchMLSchool15 The problem: We are trying to fi nd black swans A methodology that combines supervised and unsupervised learning to capture variance in data Our approach • Our objective is not to replace the reviewer, but to make her work easier • Only the reviewer is able to confidently identify real black swans • We will capture the variance in the data trying to maximize the probability of finding a black swan +3 MORE +10 MORE +6 MORE +8 MORE +5 MORE Reviewer’s feedback +2 MORE Feature extraction Clustering Anomaly Detection Supervised Models Anomalous group. It should be reviewed first Smart Video Review Framework
  • 16.
    BigML, Inc #DutchMLSchool16 Flare/Smoke Bombs Analysis Binary Classification: Smoke/Flare vs not Detection: Flare Detection Smart Video Review Framework
  • 17.
    BigML, Inc #DutchMLSchool17 Flags/Banners Analysis Detection: Waving-flag and Banner Clustering and Pattern Matching Smart Video Review Framework
  • 18.
    BigML, Inc #DutchMLSchool18 Chants Analysis raw microphones data 1 3 seconds sound blocks 2 2D spectrograms 3 Filter out high frequencies and obtain chromagram 4 Smart Video Review Framework
  • 19.
    BigML, Inc #DutchMLSchool19 Chants Analysis raw microphones data 1 3 seconds sound blocks 2 2D spectrograms 3 Filter out high frequencies and obtain chromagram 4 Chromagram Feature Extraction 5 Chants Clustering and Fingerprinting Matching 7 Chant / Noise Classification 6 Chant 1 Chant 2 Chant 3 Smart Video Review Framework
  • 20.
    BigML, Inc #DutchMLSchool20 Monkey Chants Analysis raw microphones data 1 3 seconds sound blocks 2 2D spectrograms 3 Filter out high frequencies and obtain MEL Spectrogram 4 monkey chant audio blocks non monkey chant audio blocks Monkey Chant: Periodic sound around 400 Hz BigML Image Classification model Smart Video Review Framework