of Road Images with Deep Learning Frank Geujen Senior Product - - PowerPoint PPT Presentation
Detection and Segmentation of Road Images with Deep Learning Frank Geujen Senior Product Manager William Raveane Computer Vision Engineer Mapscape, a Navinfo company GTC Europe, October 2017, Talk #23304 Who is NavInfo SD &
GTC Europe, October 2017, Talk #23304
Frank Geujen – Senior Product Manager William Raveane – Computer Vision Engineer Mapscape, a Navinfo company
NavInfo is the leading map provider in China, with focus on location big data platform, HD / SD map, Telematics and ADAS comprehensive solutions.
Automated Driving Connected Car Navigation
In Inte ternati tion
l bu busines iness exp expansion ion Adv Advanced ced te tech chnol
res esea earch rch
Mapscape Mapscape: : Com Compila pilatio tion Tech echnol
(NDS) EU EU Tech echnolo
Centre re:
Compute ter r Vis ision ion
Deep Le Learni rning
31 31 loca locali liza zati tion ba base e for r da data ta colle collect ct and d te tech chnolo
ser ervice. ice. 4 4 R&D &D Center Centers （Sh Shangh ghai、Xia Xian、Sh Shenyang、Wuhan） Beij Beijing ing Hea eadqu quarte rters rs
Ingestion/ Ingestion/ Extra xtraction tion Data Data Sour Source
10 100+ 0+
Collection vehicles
600+ 600+
dispersed field staff in China
360 360+ + cities cities
Big data mining 99 99% Highway 80 80% Main Road
22 22M+ M+
community contributions
3. 3.2+ 2+ Mill Millio ion
Signs processed per year
20 20+ Mil + Milli lion
POI updated per year
4+ 4+ Mi Mill llion ion
Road distance updated per year
Map Cre Map Creat ation ion
> 500 500
production staff
4,0 ,000 00+
page specifications
Del Deliv iver ery y
6.16+ 6.16+ Mil Milli lion n
Kilometer
24.9 .95+ Million
Core POI
260+ 260+
Attributes in the portfolio
60+ 60+
Cities of ground truth testing in 2017 Q1
80 80%/7 %/70% 0%
Update 80% POI and 70% road link in China per year field local offices
31 31
Hong Kon Kong, , Lao Laos, s, Ma Maca cao, Cam , Cambodia ia
Map data
POI OI
St Standa andard Def rd Defin init ition ion Map, Map, i is s pr prim imar arily fo ly for A r A to B to B routing
& guidance and and is is a sim a simpl plif ified represen ied representa tation tion of
the road in e road in lin links s and and nodes nodes.
Hi High gh Defini Definition ion M Map, is ap, is us used for au ed for automated/ tomated/autono autonomous
drivin ing and in and includes ludes high high accurat accurate lane e lane and and ro road ad fea features tures.
Live feed mono camera Real Time sign extraction
(Semi)
automated core map updating 2 FPS stored mono camera imagery Off line Feature extraction
(Semi)
automated core map updating Panoramic imagery Road feature extraction Projection
data Feature extraction from LIDAR HD map creation/ updating
On-board SD map Off-board SD map Off-board HD map
Technical Deep Dive:
Use case: In-car data collection on an NVIDIA Jetson TX2
Watc tch O Onli line: e: htt ttps ps://go ://goo.gl/
KBgGo8
Multip ltiple le Si Simu mult ltaneou aneous s Detect Detections
Dist istant ant Tr Traffic ic Si Sign gns
Bad ad Li Light ghting ing Cond Conditio itions
Detector: SSD Classifier: ResNet
batched data feeder for the classifier
SSD Caffe TensorRT + our CUDA kernels
Profile visualization of SSD inference 510ms 31ms
Single Image Per-Class Detection Accuracy Single Image Per-Class Classification Accuracy Single Image Detection PR Curve
Class ID Class ID Accuracy Accuracy 100 100
Technical Deep Dive:
Lane numbering Road features Gantry sign boards
Supported today:
In Development:
Crop from Panoramic Image Camera Calibration Transformation to Top View Segmentation Network Transformation to Front View Semantic Segmentation Deep Neural Network Lane Number Grouping
tch O Onli line: e: htt ttps ps://go ://goo.gl/
4CXTD5 D5
Confusion Matrix Performance of the system expected to further improve as we continue development
GTC Europe, October 2017 , Talk #23304
Frank Geujen – Senior Product Manager William Raveane – Computer Vision Engineer Mapscape, a Navinfo company