Ignacio Lpez, Rubn Salvador, Jaime Alarcn, Flix Moreno

1
Architectural design for a low cost FPGA-based
traffic signal detection system in vehicles

• Ignacio López, Rubén Salvador, Jaime Alarcón,
  Félix Moreno
• ruben.salvador_at_upm.es
• CEI - Centro de Electrónica Industrial
• ASLab - Autonomous System Laboratory
• Universidad Politécnica de Madrid (Spain)

2
Motivation

• ADAS (Advanced Driver Assistance Systems)
• Real-time analysis of complex driving scenarios
• Risk evaluation. Anticipation

3
Motivation

• Low cost systems
• Low cost CCD cameras and FPGAs
• Mass production. Introduction in the market
• Scarce HW resources
• Architecture developed for optimal use
• Blackboard Architecture
• Neural Networks in reconfigurable HW
• Real-time classification systems
• HMI (Human-Machine Interface)

4
Outline

• System Requirements
• Reference Architecture
• Blackboard architecture (BB1/AIS)
• Neural network topology
• Implementation architecture
• Hardware Architecture
• Features, optimizations, performance
• Results
• Conclusions
• System evolution

5
System Requirements
Scalability
Low cost
Flexibility
FPGA
BB1
Real time
Reconfigurability
Predictability
Integrability
6
Outline

• System Requirements
• Reference Architecture
• Blackboard architecture (BB1/AIS)
• Neural network topology
• Implementation architecture
• Hardware Architecture
• Features, optimizations, performance
• Conclusions
• System evolution

7
Reference Architecture. Blackboard Architecture

• Simplified BB1/AIS architecture

8
Reference Architecture. Neural network topology

• Neural networks ?? hugely resource consuming

super-net
atomization

• Identification of singularities
• characteristic elements of shape
• 8x8 pixel region
• Function specialized networks
• Collaborative networks
• Single layer perceptrons

• Challenge ? Performance/Cost Trade-off
• Exploit FPGA parallel architecture...
• ... and its potential for reconfigurability

9
Outline

• System Requirements
• Reference Architecture
• Blackboard architecture (BB1/AIS)
• Neural network topology
• Implementation architecture
• Hardware Architecture
• Features, optimizations, performance
• Results
• Conclusions
• System evolution

10
Hardware architecture

• Communications Interface and Preprocessor
• Minimize the use of large memory banks
• Minimize the complexity of the control

11
Hardware architecture

• Knowledge unit (KNU).
• Preliminary stimations ? impossible to implement
  more than one KNU in a single FPGA

• Problem overcome
• Build a single, generic neural network optimized
  in time to be run several times

• Datapath
• No overflows - Saturation
• Fixed point
• Signed fractional
• Q10.22 Q25.44

12
Hardware architecture

• Mode of operation
• Computational load decrease ? RoI
• Singularity detection ? RoD
• 10x9 RoDs per RoI
• RoI Extractor detects the RoI
• GNN process RoDs
• Probability detection map obtained

• Modularity
• Distributed layer control
• Layer modules
• Implementation on different FPGAs

13
Hardware architecture
Less LEs Fmax increase Scalability

• Architecture optimizations
• Binary images ? Thresholding
• Q8.16 format
• Design for reuse methodology

• GNN performance
• 73 clock cycles

14
Outline

• System Requirements
• Reference Architecture
• Blackboard architecture (BB1/AIS)
• Neural network topology
• Implementation architecture
• Hardware Architecture
• Features, optimizations, performance
• Results
• Conclusions
• System evolution

15
Results
GNN ? Single layer perceptron 30x3

• Altera CycloneEP1C20F400C6

• Altera Cyclone IIEP2C35F672C6

16
Outline

• System Requirements
• Reference Architecture
• Blackboard architecture (BB1/AIS)
• Neural network topology
• Implementation architecture
• Hardware Architecture
• Features, optimizations, performance
• Results
• Conclusions
• System evolution

17
Conclusions

• Traffic signal identification in urban areas

• Low cost

• Flexibility

• Scalability

• Integrability

• Predictability

• Reconfigurability

• Efficient usage of resources

18
Outline

• System Requirements
• Reference Architecture
• Blackboard architecture (BB1/AIS)
• Neural network topology
• Implementation architecture
• Hardware Architecture
• Features, optimizations, performance
• Results
• Conclusions
• System evolution

19
System evolution

• Fully develop Blackboard architecture
• Pedestrian detection
• System reconsideration
• GNN distribution
• FPGA
• Self adjustable RoI processing strategy
• Autonomous vehicles - Robots
• FPGA benchmark
• Cognitive architectures scalability
• Hardware embedded intelligence
• Implementation methodology
• Features conservation

20
Thank you for your attention!! ruben.salvador_at_upm.
es