Date post: | 11-Apr-2017 |
Category: |
Technology |
Upload: | sirris |
View: | 25 times |
Download: | 1 times |
1
Application specific low cost positioning for vehicles
Erwin Rademakers
Senior Research Engineer
28 March 2017
Low cost Positioning
AGENDA
Motivation & user needs
Absolute & relative positioning concepts and results
Computing platform
B2B project: Low cost and accurate positioning for cycling
WHY POSITIONING?
Persistent need for low cost accurate, reliable and robust positioning systems
Room for innovation in low cost positioning systems tailored per applicationdomains
Current price factor is 100 going from 2m to 2cm
APPLICATION REQUIREMENTS
On-road applications: Automated/Autonomous driving
➢ Sub-meter accuracies for safety, efficiency and comfort
➢ Precise HAD map through crowdsourcing
Off-road applications
➢ dm to cm range accuracies
➢ Agricultural vehicles, special purpose vehicles
POSITIONING CONCEPTS
Absolute earth coordinatesresulting from
• Satellite systems (e.g. GPS)
• Matching (land)markerswith an accurate map
Relative positioning tolimit the delta againstlast absolute coordinatesthrough Odometry
• Inertia odometry (IMU)
• Visual odometry (camera)
Fusion techniques for accurate, robust and reliable positioning system
FUSION
Positioning
Absolute Relative
MAP
INS
Odometry
Visual
Beacons
GPS
Visual
Odometry
FUSION
Positioning
Absolute Relative
Map
Visualodometry
INS
odometry
Visualbeacons
GPS
Receiver
ionosphere
troposphere
ABSOLUTE POSITIONING: GNSS SATELLITESGNSS RECEIVERS
• Single frequency - Dual frequency
• Dual frequency allows for removal of the first ordereffect of the ionosphere
• Single – Multiple concurrent constellations
• GNSS constellations of interest: GPS, Glonass, Galileo (y2020)
Error Component Description Potential Improvement
Ionosphere Free electrons (sun) influence satellite signal (EM waves), above > 1 km
200 cm
Troposphere Temperature, humidity, pressureinfluence, below < 1 km
200 cm
Ephemeris data Satellite position data 100 cm
Satellite clock drift Clock drift 150 cm
Dcb Differential code bias, hardware delay 10 cm
WindUp Rotation of the antennas 10 cm
Sagnac effect Light travels in a spiral path 5 cm
ROA Satellite orbit correction (radial, allong, crossing direction)
5 cm
Solid Earth Tides Tides: Moon – earth interaction 2 cm
Relativistic effect Relativistic clock correction 5 cm
Available solutions: GPS ≈ 25 € (10 m acc)
RTK > 5 k€ (2 cm acc)
Solution: SF-PPP* ≈ 100 € (sub-meter
accurate)
SELECTION GNSS SOLUTION
Ref ESA navipedia (modified)
NEW
In cooperation with:
• Single Frequency Precise Point Positioning based• Low cost because of single constellation (GPS) - single frequency• Error corrections from internet
BENCHMARK
Effect from non-line of sight and reflections (e.g. by trees)
Driven trajectory Error against RTK as reference
(snapshot)
Low cost GPS 25 € Low cost PPP 100 €RTK-reference > 5k €
t
error
sub
-met
er r
equ
irem
ent
RESULTS “OPEN AREA” ENVIRONMENT
No obstructions from trees or buildings.
RTKPPP
RMSe North RMSe East
Avg σ Min Max Avg σ Min Max
0.53 m 0.22 m 0.17 m 0.98 m 0.48 m 0.15 m 0.24 m 0.77 m
RESULTS “HIGHWAY” ENVIRONMENT
In this environment we have less satellites due to obstructions from trees.
RTK
Waiting @traffic light (trees)
RMSe North RMSe East
Avg σ Min Max Avg σ Min Max
1.15 m 0.30 m 0.67 m 1.73 m 0.62 m 0.28 m 0.36 m 1.25 m
WHAT IF THERE IS NO GPS OR LIMITED AVAILABLE?
FUSION
Positioning
Absolute Relative
Map
Visualodometry
INS
odometry
Visualbeacons
GPS
RELATIVE POSITIONINGINS ODOMETRY
Continuously calculate position based on MEMS IMU
• Gyroscope: degrees/s in 3D
• Accelerometer: g in 3D
• Compass (useless close to metal)
• 100 Hz
SENSOR FUSION
Kalman Based
Filtering
Sensor fusionGPS PPP
Position
Course
Velocity
IMU
Yaw rate
Acceleration
PoseYaw
Velocity
Ext. Kalman based filtering
?ˆ,ˆ ,N,E V
GPS 10 HzIMU 100 Hz
FUSION 100 Hz
DRIFT
Absolute coordinates available
Drift / Integration errorCorrection moment
to available absolute
coordinate
Errors in measurements of gyro- & accelerometer resulting in a drift.
WHAT ABOUT CORNERSIn a corner: vehicle heading ≠ position heading
Sideslip angle estimation
• Measuring wheel speed, steering angle
• Vehicle model (bicycle model, 2 wheel comparable against 4 wheel)
SENSOR FUSION
Sensor fusionGPS PPP
Position
Course
Velocity
IMU
Vehicle
Yaw rate
Acceleration
Steering angle
Wheel speeds
PoseYaw
Velocity Side slip?ˆ,ˆ ,N,E V
Ext. Kalman based filtering
RESULTS URBAN ENVIRONMENT PPP ONLY
In an urban environment we have less satellitesavailable due to obstructions from buildings.
RTK
Buildings blockingSatellite reception
Open area
RMSe N RMSe E
PPP only 3.90 m 3.24 m
RESULTS URBAN ENVIRONMENT PPP AIDED WITH IMU
PPP used in combination with IMU (gyro and accel)▪ Extended kalman filter (EKF)
RMSe N RMSe E
PPP only 3.90 m 3.24 m
PPP+IMU 1.16 m 1.14 mmore than times improved
EKF (PPP+IMU)RTKPPP
Content
HARSH ENVIRONMENT
GPS outage
EKF (PPP+IMU)
SBG
PPP
PPP large error
RMSe N E
PPP+IMU 1.08 m 0.96 m
FUSION
Positioning
Absolute Relative
Map
Visualodometry
INS
odometry
Visualbeacons
GPS
Vehicle
Camera
Road Surface
RELATIVE POSITIONINGVISUAL ODOMETRY – GROUND FACING CAMERA
Height (𝐻)
Camera
Displacement (𝐷∆𝑡)Road Surface
////////////////////////////////////////////////////
𝐶𝑡1𝐶𝑡2
Have potential 𝑚𝑚 resolution accuracy (depending on calibration, resolution …)
Using a mono camera
Finding the maximum phase correlation between two consecutive frames
A wheel encoder suffers from wheel slippage, GFC not
Independent of the kinematics of the vehicle
Closing Trajectory Length ≅ 𝟐𝟎𝒎
Error Distance = 𝟎. 𝟐𝟔𝟕𝒎 (𝟏. 𝟑𝟑 %)
Error Orientation = 𝟏𝟕. 𝟐𝟎𝟓𝟕°
EXPERIMENTAL RESULTS GFC
Data sample (Octinion)
RELATIVE POSITIONINGVISUAL ODOMETRY – STEREO CAMERA
Have potential c𝑚 resolution accuracy (depending on calibration, resolution …)
Using stereo camera
Matching visual feature points between two consecutive frames
Extracting 3D information from multiple 2D views, which provides relative depth information of a scene
• In GFC the height (depth) is already known
Independent of the kinematics of the vehicle
Matching features in a line/circle, colors encode disparities.
EXPERIMENTAL RESULTSSTEREO VISUAL ODOMETRY
Distance: 60m
Matches: 126
Inliers: 50%
Failed frames: 3
Good frames: 755
GPS
VO
FUSION
Positioning
Absolute Relative
Map
Visualodometry
INS
odometry
Visualbeacons
GPS
MAP FOR POSITIONING
Multi layer approach
• Static layer: existing environmental data
• Dynamic layer: temporal object information
• Has confidence, accuracy and reliability parameters
Positioning approach using
• Visual Beacons
• Road Signs
• Landmarks
• Potholes
• Road bumps
• ...
BOOTSTRAPPING THE MAP
DetectionVisual Beacons
Road Signs
LandmarksPotholes
Road bumps
Positioningfusion
D
relative positioning
visual odometry
Map1. Visual Beacons2. Traffic Signs3. Re-visit-able
Landmarks4. Road surface
x,y,z,rotx,roty,rotz
PPP
absolute position bootstrapping of unreliable
landmarks and road surface imperfections
FUSION
Positioning
Absolute Relative
Map
Visualodometry
INS
odometry
Visualbeacons
GPS
Content
ABSOLUTE POSITIONINGVISION BASED – PASSIVE BEACONS – ROAD SIGNS
Absolute position based on detection of road signs• Mono camera• Relative position from detection algorithm: 3D camera position (relative) from the road
sign detection algorithm and the absolute position of the speed sign.• The absolute position of the road signs are retrieved from the dynamic map
Can also be used to create a High Definition MAP• Can also be lane markers, buildings, traffic lights, obstacles (depending on the detector)• Map update via crowdsourcing to obtain an accurate position
RESULTS ROAD SIGNS
Camera resolution 1280x960 @ 20HzVelocity speed 70 – 50 – 70 – 50 kph
RMSE < 0.5 m
° Ground Truth° RS
50
70
70
50
ABSOLUTE POSITIONINGVISION BASED – ACTIVE BEACONS
Where GPS is limited available, to bootstrap the creation of a MAPControlled lighting beaconsTriangulation approach with heading fusionMotion compensationLow speed
DeltaBeacon OFF Beacon ON
Size and deformation are measures for distance
RESULTS VISUAL BEACONS
Detection accuracy related to optics & resolution
For <40m: ±25cm accurate (RMS error E/N)
For >40m: 25 – 50cm accurate (RMS error E/N)
Position Error
FUSION
Positioning
Absolute Relative
Map
Visualodometry
INS
odometry
Visualbeacons
GPS
RTK Low Cost-GPS PPP PPP/IMU Beacons Road Signs
Accuracy 1-2cm 2-10m 0,2-2m 0,2-2m < 20cm 0,1-50cm
Update rate 200Hz 1 Hz 10Hz 100Hz 20Hz 20 Hz
GFC
Road Signs
VO (stereo)
gg,V
vv,V
rrN,E
PPP/IMU
Vehicle Model
Beacons
Sen
sor
fusi
on
?ˆ,ˆ ,N,E V
V, N,E,
,,yx
aa
bbN,E
MULTI SENSOR FUSION
To obtain robust and accurate position in different situations/environments
Centralized Kalman filterTaking into account for each sensor
• Measurement noise• Weight factor
Fusion with GFC/VO status experimentalFusion with Map conceptual status
MAP
Results from real time multi sensor:• PPP/IMU – VM – Beacons – Road Signs.
Simulation of a harsh environment:• Obstruction from bridges (indoor-outdoor), trees, …
The reference systems is also suffering in this harsh environment (RTK system aided with IMU).
RESULTS MULTI SENSOR FUSION
RTK drift
RESULTS MULTI SENSOR FUSION
EKF
RTK
PPP
GPS outage
bad GPS reception
suffering from trees
PPP+IMU PPP+IMU+Road Signs
EKF
RTK
PPP
RSR
drift corrected
by road sings
PPP+IMU+Road Signs+Beacons
EKF
RTK
PPP
RSR
BEACONS
position accuracy
improved by beacons
GENERIC MULTI SENSOR FUSION
Research PlatformEmbedded platform
RESEARCH PLATFORM
+GPU
MB
Router+ 4G
SSD
CAN
Distributed architecture: open platform (extendable, flexible)
Outstanding performance using fast hardware CPU/GPU.CPU Intel 64 bit, 4x core, 3.5 GHzGPU Nvidia GTX 780
+
Trunk Ford S-MAX
MULTI SENSOR FUSION
TCP
TCP
TCP
Visual OdometryRoad Signs
Fusion
Beacons
TCP TCP
GFC
TCP
TCP
PPP
IMU
Client
>>
>>
>>
>> <<
>>
>>
>>
SOFTWARE DEVELOPMENT ENVIRONMENT
Enterprise architect for OOAD (UML)
• PlantUML in the code with doxygen
IDE eclipse CDT
• Scons (build tool)
Programming language C++11 (multi-arch, 32-64 bit), python (scripts)
Linux operating system (Ubuntu, with pre-emptive patch in development)
Main 3th party libraries (Open source)
• Boost (XML, UDP/TCP/serial communication …)
• OpenCV (computer vision, basic functions)
• Newmat (matrix operations)
• Google test/mock
Profilers: memory check, performance, code coverage, … (GNU-tools, Google)
EMBEDDED PLATFORM
Research platformCPU Intel 64 bit, 4x core, 3.5 GHz
GPU Nvidia GTX 780
Nvidia Jetson Tk1 (Tegra)CPU 2.32GHz ARM quad-core Cortex-A15 (32 bit)
GPU with 192 CUDA cores
Rpi 3CPU 1,2 GHz ARM quad core Cortex-A53 (32/64-bit)
GPU Broadcom VideoCore IV
+
3x 2x
Less
per
form
ance
COST PROTOTYPE
Embedded platform (mono camera)
Indicative costs
• HW/SW costs (depending on the use case varying from 6 MM to 12 MM)
• BOM: 300 to 400 €
Computing platform(with GPU)
Camera(low cost variant)
IMUGPSData connection
50 € 10 €50 € 150 €100 €
TEST SETUP
Platform
Ground Facing Camera
LOW COST & ACCURATE POSITIONING FOR CYCLING
Company:
Problem statement (summary):
Current products have low accuracy and low update rates.
Design of the bike computer doesn’t fit the design and “aero-dynamics” of the bike
Additional sensors required on the bike
POSITIONING FOR CYCLING
RTK
m distance
m height
kph average
kph
SUMMARY
SUMMARY
Low cost and accurate techniques can deliver high accuracy• GPS with PPP < 50 cm (open air)
• INS odometry < 2 % (error distance)
• Visual beacons < 50 cm (low-mid speed)
• Visual odometry < 1.5 % (error distance)
• Pitfalls:
• high drift in case of long outage of an absolute position (i.e. GPS outage)
• inaccuracy of vision based positioning due to light conditions, shocks, needs texture, imperfect camera calibration, dust/dirt, low vehicle speed (framerates)
BOM Cost of these techniques can stay below 400 € (prototype)
Autonomous vehicle positioning does require fusion of techniques especially in harsh environments