desk demo poster book
TRANSCRIPT
2012 2014
ADAS functions
Platform specifications
2016
Vehicle prototypes
2018 2020 2022
DESERVE DEPLOYMENT
Standardisation and marketing
Industry implementation
2nd generation tool available
Design and develop a Tool Platform for embedded ADAS
• exploiting the benefits of cross-domain software reuse
• standardising the automotive software component interfaces
• easy integration of heterogeneous modules Low cost, highly reliable, standardized Tool Platform for can seamlessly integrating different functions, sensors, actuators and HMI to enable the development of a new generation of ADAS applications.
•Methodology for the common software platform •Development of the selected ADAS functions •HMI design, driver model and driver monitoring •Verification of the tool platform methodology • Implementation of the 5 demonstrators
Budget 25 M€ EU Funding 4.2 M€ National funding 7,2 M€ Schedule Sep 2012 – Feb 2016 Programme ECSEL Joint Undertaking Coordinator Matti Kutila, VTT Partners 23 partners from 9 countries Contact [email protected] Website www.deserve-project.eu
Frank Badstübner [email protected]
DESERVE Platform:
„The journey is the reward“
ADAS application portfolio(compliant with DESERVE module concept)
Common Modules:
Lane course,
VRU detector,
Vehicle detector, ….
Functions:
Emergency brake,
VRU protection,
InterUrban Assist, ….
ADAS rapid
prototyping
framework
Specifications and
requirements for
System on Chip(SoC)
next generation embedded ADAS systems
Model Based DSEMatlab/Simulink/ADTF/RTMaps
Validation
and Test
Cost prediction
(silicon area,
throughput, Pv,…
HIL, MIL, PIL test bench
MicroAutoBox, FPGA-Board,
Embedded PC, Aurix,…
Ite
rati
ve
pro
ce
ss
> 1
5%
sp
ee
du
p
A Generic Platform Concept
Development of future ADAS systems with maximum reuse of modules and components due to well-defined processes and standardi- sations on architecture • Provides flexible framework reaching
from PC based pre-development to close-to-production HW implementation
• Constructs tool chain to allow for model- ling and evaluation via virtual testing
• Common in-vehicle platform for furture ADAS functions
• Enables the integration of safety (ISO26262) and security (ISO27001) mechanisms for pre-certification
DESERVE platform concept to speed up ADAS development process by more than 15% compared to state of the art ADAS development levels
Design and Development Process
• Development framework to seamlessly support the ADAS developments levels
• Infineon focus: level 3
model based design
software implementation
Embedded sw / programmable hw
implementation
Hardware-in-the-Loop testing
Software-in-the-Loop testing
Model-in-the-Loop testing
user acceptance
business requirements
Compilation and configuration
Driver assistance design
Function design and development
Function Configuration
Function testing & validation
is this the driver assistance we want?
does the function behave as it should?
does the function perform in real-time?
does the vehicle driver accept and use the assistance?
Level 1: PC platform
Level 2: Rapid prototyping
Level 3: Fully embedded,
AUTOSAR compatible
DESERVE platform enabled design and development process DESERVE platform enabled design and development process
Dedicated requirements for embedded multicore platform with FPGA
Development level 3 goes significantly ahead compared to levels 1 and 2 in terms of fulfilling „critical“ requirements • Autosar compatibility • SPICE compliance • Functional safety (ISO 26262, ASIL D
requires FMEDA and safety concept) • Reliability
Key achievements • FMEDA for multicore processor for use in DESERVE
development levels 3 (and 4) done • Basis elements of the safety mechanisms published
(e.g. hierarchical modularization, Safety Element out of Context – SEooC)
• Work done for „non-industrialised“ platform can be carried over to the „industrialised“ DESERVE platform
Frank Badstübner [email protected]
David González
INRIA [email protected]
Arbitration process applied to automated driving -WP24 Generic ADAS Control-
Context
An algorithm for control sharing between the driver and the automated control is presented. Once the limits for a safe driving are defined, the automated vehicle can act as a supervisor of the driver performance. Haptic feedbacks are given to the driver through the steering wheel, where the force of those feedbacks is determined by the risk assessment and the attentiveness of the driver.
Haptic Steering Wheel -communications bridge-
Arbitration and control modules
Conclusions
Our approach can effectively determine the risk of the situation in hand, evaluating the scenario and the driver status. Haptic feedbacks inform the driver of the decision of the automated systems, allowing smooth interaction between the two and enhancing safety.
Data from CRF demo vehicle
(DESERVE)
David Gonzalez
INRIA [email protected]
Path planning with obstacle avoidance and different
speed profiles
-WP24 Arbitration/Control and WP44 Automated functions- Context
A continuous curvature planning algorithm with obstacle avoidance capabilities is presented, considering different speed profiles for the improving of the comfort and reducing lateral accelerations in the driving process.
The automated system generates a collision free path that considers vehicle’s constraints, the road and different obstacles inside the horizon of view. The path planning was developed in the framework of the WP24 and WP44. The results were simulated in PROSivic and RTMaps, using real data.
Expetimental Results Conclusions
Implementing parametric curves, different intersections are handled. By searching the position of the Bezier control points, a smooth path that fits the vehicle and road constraints can be planned. The user is able to set a level of comfort in the navigation system, as this modules calculates the longitudinal speed profile.
André Rolfsmeier
Christian Lindemann
dSPACE GmbH [email protected]
DESERVE Platform Concept
ADAS Development System
ADAS Development Process
Jens Klimke [email protected]
Frederic Christen [email protected]
Context
PELOPS (PC)
Driver Model (Virtual driver)
Environmental Model (Virtual environment)
Vehicle Model (Virtual vehicle)
DESERVE platform
Perception Platform
Application Platform
IWI Platform
Control function
CA
N C
AN
Implementation
Requirements • Inter-Urban driving behaviour (incl. safe passing) • Route-motivated lane changes • Full intersection behaviour without pedestrians • Predictive driving (speed adaption, ...) • Realistic mapping of the human driving motivations • Re-usage of validated driver model approaches • As Simulink model (easy to implement, easy to debug)
Motivation Conscious Unconscious Action
Free moving Desired velocity Acceleration Pedal value
Following Desired distance Acceleration Pedal value
Lane keeping Fix-points Yaw rate Steering wheel angle
Lane change Two-step LC
Lat. offset/ Fix points
Lat. velocity / yaw rate
Steering wheel angle
Stopping Stop distance Acceleration Pedal value
Standing - Pedal value -
Safe passing Fix points Yaw rate Steering wheel angle
Concept
• Well-known structures Donges, Rasmussen • Validated driver models Wiedemann, Ehmanns, ... • Global parameterisation • Transparent data
management • Separation between
panning, manoeuvre, conscious and unconscious decisions
Navigation
Guidance
Stabilisation
Veh
icle
HM
I State
Ped
als/Steering
Drive
r
Action
Action
Action
Action
Memory State Parameters
Perception
Radio/ Global traffic information
Traffic/ road/
environment
Vehicle
Info processing
Manoeuvre decision
Conscious implementation
Unconscious implementation
Planning
Route
Manoeuvre
Reference variables
Enviro
nm
ent
Enviro
nm
ent/R
oad
/Traffic Results 1. Driver behaviour prediction
2. Simulation driver model
Context
PELOPS (PC)
Driver Model (Virtual driver)
Environmental Model (Virtual environment)
Vehicle Model (Virtual vehicle)
DESERVE platform
Perception Platform
Application Platform
IWI Platform
Control function
CA
N C
AN
Requirements • Control mechanisms with the usage of longitudinal and
lateral actuators • Demonstration of the usability and increase of
development performance with the DESERVE platform
Architecture
Prototype definition • Development close to the DESERVE hardware (here:
Matlab/Simulink model on the dSpace Micro Autobox II) • Inter-urban ACC with advanced lane keeping control
(Control of target distance, speed limits, curve speed, passing distance, passing speed, curve cutting)
• Hand-over to the driver before intersection areas • Trajectory based combined longitudinal and lateral
control
Toolchain: • MiL/HiL Simulation with PELOPS • Implementation of a realistic driver model for intersections • Implementation of Driver-ADAS-Interaction • Sensor data simulation • Controller optimisation with AVL Cameo
Jens Klimke [email protected]
Frederic Christen [email protected]
Display
Brake
Vehicle control
Target selection/
classification
IWI manager
Radar
Perception Application IWI
Frontal Object Perception
(FOP) VRU
Lane recognition (LR)
Vehicle trajectory calculation
GPS
Map
ADASIS Horizon (AH)
Powertrain
Enhanced Vehicle
Position (EVP)
Relative Positioning to the Road (RPR)
Vehicle Filter/State (VFS)
Vehicle Data
Frontal Camera
Lane course (LC)
Strategy/ Trajectory
Lane asphalt potential
Curve potential
Res. potential
Lane asphalt potential Obj. potential Res. potential
Fabio Tango – CRF [email protected]
Mauro Baldi - POLITO [email protected]
Chiara Ferrarini - ICOOR [email protected]
To develop a driver model by starting from the analysis of existing solutions. The driver model modules consists of two elements: • Virtual Driver: will be used to
faithfully simulate the driver behaviour and as a tool to test the DESERVE development platform.
• Driver Intention Detection Module (DIDM): will be used within the platform to predict driver’s intentions.
Aim The DIDM test is aimed at collecting driving data by means of car log files and Socioanagraphic & driving style questionnaires for 46 users. The test focuses in gathering data about complete 900 overtaking manoeuvres (including 2 Lane Change (LC) manoeuvre) using CRF DEMO vehicle.
DIDM test
Preliminary results were computed in Matlab with the help of the Pattern Recognition Toolbox (PRT). We trained the following classifiers on a large dataset with 3 features, binary and multiple target variables, 1035 rows for the Training Set and 450 rows for the Test Set: • Support Vector Machines (SVM) with Gaussian kernel • Relevance Vector Machines (RVM) • Hidden Markov Models (HMM). The dataset has the following features: • Speed • Steering angle • Time to collision. The effectiveness of the classifiers is expected to be > 80% Classification on data from the selected drivers is underway.
Preliminary results
Driver Model
DIDM will be used within the platform to predict driver’s intention with respect to some specific manoeuvres: • Lane-change/Overtaking • Car-following Inputs: vehicle dynamics data outputs from FOP module direction of user’s head from internal camera. Outputs: probability of the next manoeuvre the driver intends to perform and a level of confidentiality.
DIDM
Manual calibration of the SVM with RBF kernel
2 4 6 8 10
1
2
3
4
5
6
7
8
9
10
0.95
0.955
0.96
0.965
0.97
0.975
0.98
0.985
0.99
0.995
1
SomnoAlert® Sensor+ for Smartphone is an app developed to detect “not apt to drive” states using physiological signals such as thoracic effort signal.
An external thoracic effort sensor sends the respiration data to the Smartphone, where it is processed to evaluate the attention state and warn if the drowsiness state of the driver becomes dangerous
Additionally, this system works connected to a server that allows the on line control and management of a professional fleet. The fleet can use this tool to improve general safety by analyzing the drowsiness status, routes, working hours, working shifts… of the different professional fleet drivers.
•It evaluates the level of attention with physiological signals in order to achieve an objective measurement of attention •Detection of patterns to identify drowsy or inattentive states while driving. •Memory of basal thoracic effort signal in a wake state to analyse its degradation over time. •It’s car-independent.
•It can reduce accidents.
Base concept
Based on a characterization of different Thoracic Effort patterns, related with the driver state, the drowsiness level with high specificity can be estimated. The thoracic effort signal, both extracted from video signal or adquired by inductive band, is analyzed with algorithms based on the quantification of the respiratory rate variability.
Specificity 0.98
Sensitivity 0.94
Phase 0 (Alert)
Characterized by symmetry and stability in amplitude and frequency
Phase 1
(Fatigue)
Characterized by loss in amplitude and appearance of sighs and yawns
Phase 2
(Drowsy)
Characterized by high variability patterns
Benefits
Description
SomnoAlert® Sensor+ for Smartphone is an app developed to detect “not apt to drive” states using physiological signals such as thoracic effort signal.
An external thoracic effort sensor sends the respiration data to the Smartphone, where it is processed to evaluate the attention state and warn if the drowsiness state of the driver becomes dangerous
Additionally, this system works connected to a server that allows the on line control and management of a professional fleet. The fleet can use this tool to improve general safety by analyzing the drowsiness status, routes, working hours, working shifts… of the different professional fleet drivers.
Description
Detect inadequate driver 's state based
on thoracic effort information
SomnoAlert® Sensor + Detect inadequate driver 's state based
on thoracic effort information extracted from video
SomnoAlert® Contactless
HW and SW features
HW and SW features
•Route information via GPS • .It send the data to a server in order to
evaluate and help the driver •A database server storages information
sent by android application, for off-line processing
•Mobile shows information directly to the driver about their attention state
•Websites shows an historic of travelled routes and drowsiness state along the route.
The Somnoalert® Sensor+ device comprises an inductive band located around the thorax and connected to a transmitter that sends the respiratory data to the Smartphone where it is analyzed with algorithms based on the quantification of the respiratory rate variability. As a SW features the system includes:
The Somnoalert® Contactless device comprises a video camera located in
fron of the driver and recording an specific section of the thorax .
The camera is connected to a PC that analizes the video data in order
to extract the respiration signal. Once the respiration signal is extracted, it is analyzed with
algorithms based on the quantification of the respiratory
rate variability.
Adasens PAC16 camera device located in front of the driver
recording the upper body and connected to a thoracic effort signal
extraction algortihm running in RTMaps that provides the
respiration signal. Variability quantification algorithms
analyses the respiration signal and provides an estimation of the
attention degree of the driver.
Objective: Improvement of a general driver model module that will be integrated into the DESERVE platform.
The module combines a Distraction Detector and a Drowsiness Detector based on eyelids movements from partner CONTI with the drowsiness detector from partner FICOSA based on the analysis of the Driver’s respiratory rate.
FICOSA Drowsiness Detector (SomnoAlert®) comes in two different versions:
- SomnoAlert Sensor + Driver monitoring system based on a contact band thoracic effort sensor
-- SomnoAlert Contactless Driver monitoring system based on a contactless video respiration sensor
WP3: DRIVER MONITORING
Vehicle data
Audio
Display
Brake
Haptic-system
Vehicle control
Target selection
Threat assessment
Driver intention - overtaking/ lane change
IWI manager
Front camera (LDW) CONTI
Impairment detection algorithm
Eyelid motion CONTI
Drowsiness detection by
respiration rate FICOSA
Eye gaze Head pose
CONTI
Image respiration
extraction(opt) FICOSA
Bluetooth communication
FICOSA
Drowsiness detection through
eyelid motion CONTI
Driver im-pairment diagnostic
Drowsiness detection
fusion
Driver Monitoring
Inductive band
FICOSA
Interior Camera (thorax) FICOSA
Distraction detection
CONTI
Perception platform Application platform IWI platform
Interior Camera Simulations Inductive band Demo vehicle
Interior Camera
(driver face) CONTI
Andrea Saccagno P.M. Research - ADAS FICOMIRRORS ITALIA S.r.l. C.so Cuneo, 15 10078 Venaria Reale (TO) - Italy [email protected] Tel. +39-011- 0130165 -
Focus group participants expressed their opinion on the proposed DESERVE HMI concepts evaluating the strategy for drowsiness alert (explicit vs implicit).
The 2 HMI concepts identified were tested in a driving simulator with 30 users to identify the preferred one. Furthermore 9 icons to express DROWSINESS concept have been tested with 63 users.
HMI CONCEPT 2 – IMMERSIVE HMI The interaction is distributed along the dashboard and the windscreen.
HMI CONCEPT 1 – HOLISTIC HMI All the HMI elements are centralized in front of the driver. The cluster is the main visual output channel, while the steering wheel is the main input channel.
Elena Maria Bianco CRF [email protected]
Chiara Ferrarini ICOOR [email protected]
Elisa Landini RE:LAB [email protected]
Eva M. García Quinteiro CTAG [email protected]
.
The HMI FINAL CONCEPT with explicit drowsiness merges winning features from concept 1 – unique display and 2 - central warning info.
HMI CONCEPT 1 with explicit drowsiness
HMI CONCEPT 2 with explicit drowsiness
HMI CONCEPT 3 – SMART HMI The dashboard display is replaced with a nomadic device (e.g. smartphone/ tablet).
Dr.-Ing. Martin Kunert
Robert Bosch GmbH [email protected]
Modular FPGA-based, embedded Radar Prototyping Platform
Acronyms: FDM-Frequency Division Multiplexing OFDM-Orthogonal FDM MIMO-Multiple Input Multiple Output CFAR-Constant False Alarm Rate RCS-Radar Cross Section (★) source: Prof. Blume | Institute for Microelectronic Systems, Hannover (Germany)
Objectives and Task Description Methods and Steps Results and Examples
Higher sensor raw-data rates up to tens
of Gbit/s are new aspects that can’t be
handled on PC-based development
systems
Real-time signal processing on a
powerful and flexible prototyping
framework with high percentage of
transfer to the final target hardware is
required for future radars
Both low-level, massive parallel number
crunching work and very complex, more
stream-oriented and often branching
high-level tasks require application
specific, mixed processing architectures
Cost model based design space
exploration (DSE) methods will enable
the definition of the final hardware
architecture in a very early development
stage
Future automotive radar systems need
much higher angular resolution,
realized by
Higher speed resolution
Range-Doppler
processing
Novel modulation
concepts like fast-chirp,
FDM, OFDM
New antenna concepts
like
MIMO antennas
Conformal antennas
In MIMO chirp-sequence radars data
rates turn out to become extremely
demanding
With the flexible and powerful DESERVE
development framework this challenge
can be solved
An operative MIMO radar prototype
frame-work is operational at Bosch
within the DESERVE project
High-resolution range-Doppler
measurements are no w possible to
enable radar based road users
classification
Fig. 1 Trend towards heterogeneous HW platforms (★)
Fig. 2 Design space exploration for digital signal processing (★) Fig. 5 Block diagram of a three-phase MIMO radar framework
Fig. 4 Data rates w.r.t. the key radar system parameters
Fig. 3 Lateral speed requirements
Fig. 6 The DESERVE MIMO Radar prototypes
Fig. 7 Range-Doppler measurement of different road users
Fig. 8 Fast chirp MIMO radar signal data evaluation chain
Fig. 9 Wheel motion detection of an approaching bicycle
8TX/16 RX Frontend
with MATLAB interface
4TX/ 8 RX Frontend
with real-time FPGA
signal processing
both wheels
are rotating
rear wheel
is locked
both wheels
are rotating
T1 T2 T3
Range-Doppler Spectrum CFAR Threshold Peak Selection Angle determinationReal World Scenario
2 4 6 8 10 12
Range [m]
4
3
2
1
0
-1
Velo
city [
m/s
]
2 4 6 8 10 12
Range [m]
4
3
2
1
0
-1
Velo
city [
m/s
]
2 4 6 8 10 12
Range [m]
4
3
2
1
0
-1
Velo
city [
m/s
]
0
-10
-20
-30
-40
-50
RCS [dBsqm]
Dipl.-Ing. Frank Meinl
Robert Bosch GmbH [email protected]
Main objectiveAssist the driver on connection roads between cities (highways, country roads, rural roads), especially during the night.
The Inter-Urban Assist uses state-of-the-art digital maps, radar and image processing to provide advanced vision enhancement.
It incorporates modules already in series production and – additionally – newly developed functionalities.
Functionalities
Inter-Urban Assist
IUA Demonstrator
Augmented night vision system
N. von [email protected]
Mercedes Benz S-Class.
Demonstrator vehicle
Module architecture
Radar
Front
Cameras
Digital Map
Lane Course
Scene
Labeling
3D
Reconstructio
n
Self
Calibration
Vehicle Data
raw
raw
calib
data
labeled
image
trajectory
spatial
structure
feedback
ADASIS
Road course shown in
night view display
� Driver guidance on
curvy rural roadsRealized by localization in digital
map, based on Scene Labeling
Pedestrians overlaid with
distance colouration
� Driver warning for
immediate perceptionClassified by Scene Labeling,
reconstructed with Stereo
High beam turns according to road course
� Perfect illumination on curvy rural roadsRealized by localization in digital map, based on Scene Labeling
The Inter-Urban Assist is a complex
multi-platform system, realized in a
highly heterogeneous architecture.
Well-defined interfaces and the ADTF
message bus ensure the smooth
communication between modules,
across multiple instances of ADTF
running on multiple PCs.
ADASIS
LR
ARS310
NV3_CAM1
NV3_CAM2
VEHICLE
COMBI
HLC
NV3-BoB
ARS310-BoB
SYNCBOX
IMG_HUB_CUT
MAP-PC
MPC
PC_INTFCE
MAP_INTFCE
RADAR_INTFCE
VEH_INTFCE
IMG_INTFCE
HLC_INTFCE
DISP_INTFCE
LC1_SP LC2_UM
LC3_MM
LC4_EM
LC5_PC
AUGM
3D-R1_R
SGM_INTFCE
PC1
PC2
3D-R2_SM
SGM-FPGA
SL1_PP
SL2_NR
SL3_CMLGC
SC1_FE
SC2_FM
SC3_BA
FPGA
Predictive front light
Dr. R. [email protected]
Dr. L. Krü[email protected]
Main objective Scene Labeling is used to detect pixel regions in cameras images. It is used by the augmented night vision system to detect pedestrians and the road for lane course estimation. Scene Labeling is achieved by a per pixel classification using a Multiscale Convolutional Neural Network, which is trained on several thousands hand labeled images in an offline process.
Scene Labeling Workflow based on Deep Learning
Scene Labeling for the IUA
M. Limmer [email protected]
Input from camera
Image Pyramid construction
Application of the multiscale CNN Class membership probability maps Pixel classification Preprocessing
Upsampling and Classification
Local Normalization
Convolution Layer
Pooling Layer
Applications
Implementations
Lane course estimation Pixels classified as road are fitted into a relative localization model to estimate the close range lane course
Object detection Coherently classified pixel regions are segmented into objects, tracked and highlighted.
FPGA Implementation Efficient implementation on a Virtex 6 based ML605 evaluation board from Xilinx (IMS Leibniz Universität Hannover)
GPU Implementation High speed implementation for the CUDA based GPU nVidia GTX Titan (driveU Ulm University)
N. von Egloffstein [email protected]
Decomposition of the multiscale CNN Further identification and decomposition of the above mentioned Convolution, Pooling, Upsampling and Classification Layers into generic building blocks was performed. These are used to implement a hardware realization after performing a cost evaluation and design space exploration based on quantitative cost models. (IMS Leibniz Universität Hannover)
Dr. R. Schweiger [email protected]
Connection of ADTF and FPGA Emulation Board Model-Based Design Space Exploration
Dipl.-Ing. Nico Mentzer
Institute of Microelectronic Systems
Prof. Dr.-Ing. Holger Blume
Institute of Microelectronic Systems
Jun.-Prof. Dr.-Ing. Guillermo Payá Vayá
Institute of Microelectronic Systems
Institute of Microelectronic Systems
M. Sc. Florian Giesemann
Institute of Microelectronic Systems
PC
ETH
ADTF
Filter Filter Filter
ETH
MEM
FPGA Board
FPGA
PE PE
PE
MemCtrl
PE
Sensor
Altera DE2 Xilinx ML605 Altera PCIe 385N Xilinx VC707
Youtube video[2]
Hardware in the Loop [1]
Continuous improvements in advanced driver assistance systems (ADAS) require
increasingly sophisticated multimedia algorithms, which rely on enormous
processing power
A high innovation rate in automotive applications leads to to rapid changes and
many different ADAS algorithms
State-of-the-art software environments, e.g., ADTF, facilitate ADAS implementation
The development process for automotive applications and the HW/SW-co-design
process can be accelerated by implementing complex algorithms on FPGA-based
rapid prototyping platforms
Integration of FPGA emulation platforms in ADTF software development framework
increases simulation speed for algorithms and thereby supports the development
process
ADTF filter with underlying FPGA board (orange filter)
ADTF software development framework with case study application
Connection realized via Ethernet
ADTF filter communicates with
emulation board
Ethernet module in Hardware
Communication library for
software
Underlying framework
supports different
boards and is extensible
algorithmic constraints
architectural constraints
device constraints
technology constraints
Model Library
quantitative cost models
algorithmic characteristics
architectural characteristics
device characteristics
technology characteristics
intrinsic requests
extrinsic requests
intrinsic characteristics
extrinsic characteristics
For a rapid evaluation of hardware platforms by comparing different realizations and
to find the best possible implementations with fixed resources, a design space
exploration is necessary
The evaluation of hardware platforms and derivation of associated hardware costs is
a laborious and time consuming task
To shorten the design space exploration, quantitative cost models of hardware
modules can be employed
Results are intrinsic and extrinsic characteristics of a possible design, which lead
to a reduced design space
Intrinsic requests
Properties, which are predetermined
by the application itself
Algorithmic contraints
Architectural contraints
Extrinsic requests
Properties which are given by the used
device and the employed technology
node
Case Study – Cost Models for SIFT Feature Extraction
Support of multiple FPGA Evaluation Boards
MCPA board BEEcube BEE4
Compose application from building blocks
Model component with several model functions
0
50
100
150
0 2.000 4.000 6.000 8.000 10.000
Design space for LUTs Exemplary design space for the number of LUTs related to the execution cycles. The design specifications area indicated by the orange lines. The possible hardware setups are highlighted by the orange box. The theoretical design optimum is depicted by the purple circle.
design opt.
[2] https://www.youtube.com/watch?v=pX-fFX7kr5M
#exec. cycles 106
103 #LUTs
[1] F. Giesemann, G. Paya Vaya, H. Blume, M. Limmer, W. Ritter: A Comprehensive ASIC/FPGA Prototyping Environment for Exploring Embedded Processing Systems for Advanced Driver Assistance Applications, SAMOS 2014
FPGA-Based Extraction and Matching of Image Features
Application and Algorithm SW/HW-Codesign ADTF Integration and Results
Prof. Dr.-Ing. Holger Blume
Institute of Microelectronic Systems
Jun.-Prof. Dr.-Ing. Guillermo Payá Vayá
Institute of Microelectronic Systems
Dipl.-Ing. Nico Mentzer
Institute of Microelectronic Systems
Institute of Microelectronic Systems
Performance
SIFT SW-mode ~ 1 fps @ 95 W SIFT HW-mode ~ 7 fps @ <4 W Matching SW-mode > 1.000 fps @ 95 W
FPGA Ressources* (build. of pyramids, loc. of feature points, comp. of gradients)
Wide-Baseline Camera Systems in Vehicles [1][2]
The motion and vibration of a vehicle leads to
decalibration of wide-baseline camera systems
State-of-the-art algorithms for the calibration of wide-
baseline camera systems rely on robust sets of
corresponding 2D image points of the same unknown
3D points
Computation of sparse pixel correspondence maps by
extraction and matching of robust SIFT-image features for
online camera calibration
Building of gaussian scale
space
Location of feature points
in the scale space
Assignment of orientation
based on local image
gradient directions
Computation of keypoint
descriptor based on local
image region around
keypoint positions
Extraction of Image Features
Extraction of Image Features
Geometry Based Feature Matching
Online Camera Calibration
Sparse Pixel Correspondences
SIFT Feature Extraction
Building of Pyramid
Locating of Feature Points
Assignment of Orientation
Computation of Descriptor
SIFT-Algorithm regular arithmetic
control complexity
ML605 Evaluation Board with Xilinx Virtex6 FPGA
Heterogenous SIFT-System
FPGA Building of Pyramid
Locating of Feature Points
Computation of Gradient Angle
and Gradient Magnitude
ETH
Processor
Refinement of Feature Points
Assignment of Orientation
Computation of Descriptor
ETH
Geometry Based Feature Matching
SIFT-Filter
SW/HW-Mode
Adjust SIFT-parameters for algorithmic
examination
Matching-Filter
Choose out of 5 different matchers
Adjust matching threshold for algorithmic
examination
Adjust size of search window for algorithmic
examination
Visualization
Display execution times of feature extraction &
matching
Choose between 3 different visualization modes
ADTF Filter
#Reg 19,167 6.3% #LUT 26,995 18.6% #RAMB8 162 19,5% #DSP48 176 23.0% 100 MHz system clock
*Xilinx Virtex-6 XC6VLX240T, PAR-report
Improvement of quality for feature matching
Global matching Geometry based matching
right input image left input image
Overlay of input images with correct and false
matches
Color code: yellow – correct | blue – false matches
Significant increase of matching quality leads to
robust postprocessing of pixel correspondences
Computational intensive tasks on FPGA
Control intensive task on processor
Accelerated computing of SIFT features by
exploiting plattform specific advantages
DESERVE Platform
Flexible DESERVE platform leads to
interchangeability of functional modules
Rapid prototyping approach supported by FPGAs
ensures sufficient processing power for state-of-
the-art and future ADAS algorithms
[1] N. Mentzer, G. Payá Vayá, H. Blume: Analyzing the Performance-Hardware Trade-off of an ASIP-based SIFT Feature Extraction, Journal of Signal Processing Systems 2015
[2] N. Mentzer, G. Payá Vayá, H. Blume, N. von Egloffstein, W. Ritter: Instruction-Set Extension for an ASIP-based SIFT Feature Extraction, SAMOS 2014
2nd set of features
FPGA-Based Scene Labeling using Neural Networks
Scene Labeling Soft-Core Implementation ADTF Integration and Results
M. Sc. Florian Giesemann Institute of Microelectronic Systems [email protected]
Jun.-Prof. Dr.-Ing. Guillermo Payá Vayá Institute of Microelectronic Systems [email protected]
Prof. Dr.-Ing. Holger Blume Institute of Microelectronic Systems [email protected]
Institute of Microelectronic Systems
Scene Labeling Classification of every pixel in a scene Object/Obstacle detection Road and Road-lane detection Night-vision augmentation
conv
conv
conv
dataS
dataM
dataL
dataS
dataM
dataL
pool
pool
pool
dataSdataS
dataSdataS
dataSdataS
conv
conv
conv
dataSdataS
dataSdataS
dataSdataS
pool
pool
pool
fc
Convolutional Neural Networks State-of-the-Art classifier Layers of different operations
Convolution Pooling Squashing
Computation intensive 6.8×109 operations per frame
(image size 1024×512 pixels)
Instruction
MemoryPC
Instruction
Decoder
Multi-Shared Register File
64 registersX2
MO
DE
Issue 0
DMA
Controller
Write b
ack p
ath
s
IF/DE
Instruction
Decoder
Issue 1
X2
MO
DE
DE/RA
RA/EX1
RFUALUData
Memory
EX1/EX2
EX2/WB
Forw
ard
ing p
ath
s
Shared
Memory
On-Chip
Bus
On-the-fly
computations
Reconfigurable interconnection port matrix
RFU
Reconfigurable
co-processor
unit
MAC
Attached
reconfigurable
processing unit
Forwarding
Convolution
Engine
TUKUTURI1 Soft-Core Processor Generic Architecture template VLIW: Two issue slots SIMD: 64-bit registers
(1×64, 2×32, 4×16, 8×8) DMA for data transfer between external &
internal memory
ADTF-FPGA-Coupling Connection of ADTF filters with FPGA
development board Transfer filter input data to FPGA-board Computing intensive tasks implemented in
hardware
Implementation Results Speed-up with instruction-set extension ~5.1 Speed-up due to memory optimizations ~2.4 Current framerate 0.99 fps
Example of Labeled Image
Application-Specific Extensions Co-processor for convolution Programmable filter coefficients Computes 16 filters in parallel 4x16 subwords parallel Pipelined: One result per cycle
Extended DMA Controller Block transfers Queued transfers
Neural Network Topology
TUKUTURI Architecture and Instruction-Set Extension
ADTF Project for Scene Labeling
Performance of Hardware Implementation
0
0,2
0,4
0,6
0,8
1
1,2
Basic Conv. Eng. Mem. optimized
fps Framerate
[1] Payá Vayá, G.; Burg, R. & Blume, H.: Dynamic Data-Path Self-
Reconfiguration of a VLIW-SIMD Soft-Processor Architecture. Workshop on Self-Awareness in Reconf. Comp. Systems (SRCS) Int. Conf. on Field Programmable Logic and Applications (FPL 2012), 2012.
Giesemann, F.; Payá-Vayá, G.; Blume, H.; Limmer, M.; Ritter, W.: Deep Learning for Advanced Driver Assistance Systems, in: Towards a common software/hardware methodolgy for future advanced driver assistance systems – the DESERVE approach. Book to be published.
Joshué Pérez Rastelli
INRIA [email protected]
Trajectory Generator for urban intersections
-WP44 Automated functions- Context
An algorithm for dynamic path generation in urban environments is presented, taking into account structural and sudden changes in straight and bend segments.
The results present some improvements in path generation (previously hand plotted) considering parametric equations and continuous-curvature algorithms, which guarantees a comfortable lateral acceleration. This work is focused on smooth and safe path generation using road and obstacle detection information.
Segmentation process in intersections
Expetimental Results Conclusions
An Intelligent Trajectory Generator, which achieves a real time path generation engaging and safety planning to the specific characteristics of the vehicle and the intersections is presented. The longitudinal speed was adapted to keep the comfort acceleration in the experiments.
Abhishek Ravi
AVL LIST GMBH [email protected]
Model based Tuning of Map Adaptive ACC Function Velocity control to safely enter/exit
curves with comfort and in optimum manoeuvre time
Parameters Tuned: Max. Acceleration, Start of Deceleration and Control Gain
Dr. Matti Kutila Tel. +358 40 820 8334
Email: [email protected]
VTT test vehicle: Fiat 500L
Current DESERVE functions - Blind spot detection (VRU Safety) - Driver distraction warning - Low speed collision avoidance Also EU projects CoMoSef, VRUITS, and TEAM utilised the vehicle in co-operative function tests
Environment perception: - 1 Bosch LLR2 (16/4, 200m)
- 1 IBEO Lux Lidar (110/3.2, 200m)
- 1 FLIR PathFinder 320x240 (24/18)
- 1 Vislab 3DV-E29 640x480 (72/54, 65m)
- 2 Continental SRR 208-2 (150/12, 50m)
Driver monitoring: - 2 full-HD autofocus cameras
Driver HMI: - Display
Devices
Backoffice
GPS
RTK-GPSBase station
HMI
V2V & V2I co-operative
functions 802.11p
4G (LTE)
WLA
N
Stereo
camera
Thermal camera
Radar 77GHz
Laser scanner
HMI No pedestrians nor obstacles
Perception working OK
All obstacle arrows OFF
Pedestrian(s) and obstacle(s)
Pedestrian warning
Relevant arrow(s) ON
Only obstacle(s)
Relevant arrow(s) ON
General warning
Erik Nordin
Volvo Technology [email protected]
Volvo driving simulator: Volvo FH 6x2
DESERVE functions - Adaptive Cruise Control (ACC with S&G) - Collission Warning and Emergency Braking
(CWEB)
Context PreScan
DESERVE Platform
Driver Model Environmental
Model
Volvo FH Vehicle Model (Chassi and powertrain)
Perception Platform
Application Platform
IWI Platform
Erik Nordin
Volvo Technology [email protected]
Volvo concept truck: Volvo FH 6x2
DESERVE functions - Adaptive Cruise Control (ACC with S&G) - Collission Warning and Emergency Braking
(CWEB)
Platform Integrated
Mr. Aarno Lybeck Tel. +358 44 714 3763
Email: [email protected]
TTS concept vehicle: Iveco Stralis 560 e6
Current DESERVE functions - Blind spot detection - Driver distraction warning - Low speed collision avoidance DESERVE platform has been used in developing these functions for training professional truck drivers
HMI No obstacles, system working OK: Obstacle(s) or pedestrian(s):
Perception working OK
360° view, no side lights
All obstacle arrows OFF
Detections (person&pole) and side lights
General warning
Relevant arrow(s) and corresponding side light(s) ON
Obstacle illumination is part of the HMI; enlighted obstacles are more obvious in the driver display - especially in night-time. Driver gaze monitoring is utilised e.g. in factual feedback for driver candidates.
Environment perception: - 3 Vislab 3DV-E29 640x480 - 5 Continental SRR 208-2 - 8 Ultrasonic sensors
Driver monitoring: - 3 full-HD autofocus cameras
Driver HMI: - Obstacle warning screen - ASL 360 camera system screen - Iveconnect integrated screen
Devices
Nereo Pallaro
C.R.F. [email protected]
CRF concept car: FIAT 5ooL
CRASH
No Danger Pedestrian in danger detected Visual Warning + PreFill
(without distraction) No Driver reaction Autonomous brake
Function Autonomous Emergency Braking (AEB) for Pedestrian with integrated Driver Distraction
Enhanced Warning Strategies With integrated Driver Distraction
DESERVE platform From MIL/SIL simulation to on vehicle RCP
Audio-Visual Warning + PreFill
(with distraction)
Functional block diagram System architecture
C-CAN
B-CAN
LDW PRIVATE CAN
LOCAL CAN
Peak PCAN router
VGA
LDW CAMERA (CONTI)
INTERIOR CAMERA (CONTI)
MRR RADAR (SMS)
USB
DEMO DISPLAY
RS232
dSPACE
DIGITAL
Vehicle data
Audio
Display
Powertrain
Brake
Radar
Front
Camera
Interior
Camera
Driver
Distraction
APPLICATION PLATFORM IWI PLATFORMPERCEPTION PLATFORM
Vehicle
control
Threat
assessment
IWI manager
Obstacle
selectionVehicle
trajectory
Warning icons by CTAG
Toolchain for the integrated development & validation of ADAS functions
HIL – Step 1
Embedded Controller
Embedded PC
Mid-Size Vehicle
Dynamic Simulator
Perception Platform Application/IWI Platform
Prescan, ASM RTMaps Matlab/Simulink
SIM
ULA
TIO
N
TOO
LCH
AIN
PC PC
Mid-Size Embedded PC Embedded Controller
PC
Environment & vehicle dynamics
Function Autonomous Emergency
Braking (AEB) for Pedestrian.
Simple scenario
Complex scenario
Paolo Denti C.R.F. [email protected]
HIL – Step 2 MIL