6. Viola-Jones Algorithm
Binary degenerate decision tree
Feature check 1 Feature check 2
Not a face Not a face
Feature check N
Feature check 3
Not a face
~6,000
https://generated.photos/faces
16. Which features are useful and how do we extract them?
https://generated.photos/humans | https://engineeringlearn.com/10-types-of-construction-vehicles-and-their-uses-with-pictures-names/
17. Which features are useful and how do we extract them?
Let’s just throw everything at it and let the computer decide!
Lots of kernels
27. 195
Deaths
107,355
Traumatic injuries
$62b
Societal cost
There’s a moral and economic imperative to improve
heavy industry safety
2015 Australian workplace safety data
65%
Caused by mobile plant,
transport, and machinery
1,000x
Near misses
10,000x
Unsafe behaviour and
hazards
28. How can we improve heavy industry safety?
Hierarchy of
controls
Policies and procedures rather than
solution
Passive user devices
(e.g. reversing cameras, mirrors)
Users must be paying attention
Proximity
Cannot differentiate objects and issues
with cluttered environments
Tag/beacon
Every object must be tagged
AI vision
Only recently possible
33. BLINDSIGHT
INDEX
The new lead
safety metric
Does not rely on an accident
Does not rely on manual reporting
Focuses on the high-risk scenarios
Can be compared across machine types,
site, company, global
Number of near-misses
Machine danger hours
37. Impact on safety statistics
What is the impact on renumeration/bonuses for employees with safety KPIs?
Roles and responsibilities
Who should review the data?
Processes
Can all near-misses be investigated and signed off by the CEO?
Insurance
What is the impact on insurance premiums?
What are the implications of a new standard?