The document discusses autonomous vehicle research at the Surrey Autonomous Testbed. It provides details on the testbed vehicle, sensors used including cameras and lidar, and supporting research projects. The testbed aims to build and demonstrate a fully autonomous vehicle capable of parking and summon functions for under £15K. It also briefly outlines different waves of artificial intelligence including handcrafted knowledge, statistical learning, and contextual adaptation as it relates to autonomous driving work.
4. Surrey Autonomous Testbed
• The testbed is very similar to the street drone, using the same Twizy vehicle,
sensors and processing platform
• Twizy Cargo
• Nvidia PX2 Drive
• 6 Sekonix cameras (street drone as 8 but same coverage, less overlap)
• Ouster 64 beam lidar
• Wheel odometry, not available on street drone
• Supports wider research and collaborative
projects with Parkopedia, JLR, McLarren,
FLIR amongst others
• Provides Surrey with a data
capture and demonstration platform
5. Surrey Autonomous Testbed
• Two phase build:
• Phase 1 – sensor recording and onboard processing
• Phase 2 – actuators to provide full control
• Entire build opensource:
http://autonomous.home.blog
How to build an autonomous vehicle for
< £15K
7. Demonstration scenario
1. Parking process is
started via
smartphone
2. Vehicle enters a car
park, localises itself
using indoor maps
and navigates to a
space
3. Driver summons
parked vehicle back
via smartphone
4. Vehicle exits a car
park and is picked by
a driver in a pick up
zone
8. Work related to Autonomous Driving
• Automatic extrinsic calibration of car
mounted sensors
• Reinforcement Learning for vehicle
control and safety verification
• Feature Learning
across lighting and
weather
• Vehicle state estimation,
sensor fusion and side slip
estimation
• Semantic Segmentation
• Monoscopic depth estimation and
scene estimation
9. DARPA 3 Waves of AI
• Wave 1 – Handcrafted Knowledge
• e.g. classical AI, chess computers
• Engineers create sets of rules to represent knowledge in well‐defined domains
• The structure of the knowledge is defined by humans
• The specifics are explored by the machine
• Enables reasoning over narrowly defined problems
• No learning capability and poor handling of uncertainty
• Wave 2 – Statistical Learning
• e.g. machine and deep learning, alpha go
• Engineers create statistical models for specific
problem domains and train them on big data
• Nuanced classification and prediction capabilities
• No contextual capability and minimal reasoning ability
• Statistically impressive, but individually unreliable
• Wave 3 - Contextual adaptation
• Systems construct contextual explanatory models for classes of real world phenomena