#StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
Viktor Sdobnikov - Computer Vision for Advanced Driver Assistance Systems (ADAS): From R&D to Production
1. Computer Vision for Advanced Driver Assistance Systems (ADAS)
From R&D to Production
Viktor Sdobnikov
EECVC June 2016
2. Introduction
2
COMPUTER
VISION
Pattern
Recognition
Image Processing
Physics
Signal Processing
Artificial
Intelligence
Mathematics
• 1957 (Rosenblatt, Perceptron) –
probably, Birthday
• Took results and approaches
from statistics, discrete
mathematics and other fields but
at some point started to return
fundamental results: statistical
learning and unsupervised
learning theory, two dimensional
Chomsky grammars
generalization etc.
• Heuristics and invalid practical
approaches, important at the
beginning, brake the evolution
today
3. ADAS use cases
3
City Driving
Pattern
Active Park Search
Advanced ACC
Augmented
Navigation
Info-graphics
Help in low visibility mode
Recognition based
dynamic DB update
4. Automotive limitations and restrictions
4
• Intel Celeron
• 1296MHz - 1024 Mb RAM
• 256 Kb L2 Cache
• Resources are partially utilized by navigation,
voice recognition, RSE data exchange, etc.
• Real-time system has strict requirements on
latency
• System should be “up-to date” during 6-15 years
9. Proper problem formulation
9
Examples:
• w(d,d')=0 if d=d', 1 if d!=d‘
• eps if d=unknown
• w(k,k')=0 if |k-k‘|<d, 1 if |k-k‘|>=d
• w(k,k')=|k-k‘|
• No a priory probability
• Lack of training data (heart disease
vs skin color segmentation)
10. Proper and improper strategies - example
10
https://drive.google.com/file/d/0B3h0DfKqcKP9M01FTW1ic09tdTg/view
12. From R&D to production: Criteria and Acceptance
12
Case sensitive:
a. Roundabout, lack of markings
b. Highway, moderate traffic
13. From R&D to production: Testing
13
HW and SW based approach of
data collection, storage and
usage in test suite for testing
on different levels of various
ADAS applications