How AI is Disrupting
Traffic Management in Smart City
Jorge Sebastiao, CISSP
CTO Ecosystem
DCSS
Twitter.com/4jorge
Linkedin.com/in/sebastiao
Human Society Is Entering the Intelligent Era
Data
Generated
Data
Stored
Data
Processing
Actionable
Insights
Big Data Artificial Intelligence
Structured and unstructured
data
Data is stored in databases
and servers
Process the data using
CPU/GPUs and AI algorithms
to detect patterns
Predictive signals are
generated
ICT infrastructures have become the cornerstone of the smart world.
Big Data + AI  EI
Intelligent analysis
AI+ Database +Hadoop
Big data analysis
Hadoop Fusion analysis
Hadoop + Database
2013
2015
2017
Evolution towards Enterprise Intelligence
Theoretical Basis Internal Practice
Chip Device Cloud
External Enabler
(EI services)
Noah Ark
Lab
Distributed
Lab
Neumann
Lab
Shannon
Lab
GTS ISC+
Device BP&IT
Heterogeneous computing
General AI
Essential platforms
Domain-specific
solutions
Motivation: Connected Intelligence in Enterprises
Collocated
storage
Device-Cloud
collaboration
HW-SW
synergy
Easy to
Use
Converged
decision
Full lifecycle
support
Open &
compatibility
EI Platform Services
Easy & Inclusive: Put Enterprise Intelligence Within Reach
Feature ranking
Feature
engineering
Model tuning
AutoFE & Model Optimization
Full-stack AI capability
Big data base
Intelligent Decision
Essential Platforms
Generic AI Services
Rich services ready to use
Full Lifecycle Support
Train
Evaluate
Serving
Deploy
Upgrade
Intelligent Computing Platform elements
AI Accelerator Module
AI Appliance
AI Accelerator Card AI Edge Station
Heterogeneous Server Intelligent Video Analytics Server
GPU FPGANPU
Pre-integrated solutions for Smart Cities
Intelligent Twins
 TrafficGo: Traffic optimization, safety, trend
analysis
 iWater: Sewage treatment, environmental
monitoring, river safety
City
 Production quality:Flexible quality classification,
predictive maintenance
 Sales forecast: Regional, product, or combined
 Production planning: Material forecast
Manufacturing
 Customer analysis, customer profile, customer
forecast, VIP analysis
 Abnormal event monitoring: Traffic surge
monitoring, illegal behavior (stay/wander)
Campus
 Inventory management, shipping
path planning
 Trunk dispatching, smart entrucking
 Intelligent declaration: OCR,
customs price forecast
Logistics
 Video recommendation: user
profile, video tag, video classification
 Content inspection:
Pornography, terrorism, violence...
Internet
 Fleet management: Electric fence, trajectory
mining, path planning
 Autonomous Driving: MDC, V2X, Scheduling
Vehicle
 Home Safety: Stranger detection
 Family time: Smile moment, pet identification
 Elderly care: Fall detection, cry detection, abnormal
sounds
Home
 Pathological analysis: Cervical cancer analysis, lung
cancer tissue analysis, pathology report generation
 Genetic analysis: Gene sequencing, disease
classification and prediction based on genetics
Health
Converged : Crowd Intelligent Decision & Planning
city
district
intersection
Hierarchical & Crowd
Intelligent Decision
lights
truck
camera RLS
MLSGES
Converge
DLS
Smart Transportation
Supply Chain
Computer Vision
Urban Transportation Challenges
---Congestion, Violation, Management
• Violation Types: Illegal Parking、
Improper Lane Use、Crossing
Line(Driving) and Wrong Way(Driving)
• Violation Number: ~300k Case per
Day in X city
• Large Volume of Vehicle:~6
Millions Vehicles in X city.
• Congestion:The average speed is
~22km/hour in the rush hour。
Severe Congestion Frequent Traffic Violation Low efficiency Management
• Workload: ~300k image capture
per Day, but only process 30k-50k
due to workload
• Scope: Lots of Blind Spot. Non-
AI camera.
Camera Video
Signal Control Policy
AI Based Signal Control
Offline Variant Cycle Online Adaptive
Video Data Analysis
Deep Learning
Traffic Data Fusion
Huawei AI Platform
pedestrian recognition Vehicle recognition
Edge Computing Cloud Computing
Object Tracking
Deep Learning
Signal Control: Integrate signal control theory with
reinforcement learning to support big area signal control.
Edge Computing:Great cost performance with Huawei AI
chip
Video Analysis:Analyze the pedestrian, Vehicle and Road
Structure by deep learning; bird's-eye view video
Concise Traffic Data: Fuse the structured data from
traditional device with video.
Edge Computing
Reducing Congestion by Signal Control &Traffic Situation Perception
Structured Data
from Other
Device
Lane recognition Tiny Object recognition
Image Enhancement
Passenger
Car Unit
Diverging
Ratio
Vehicle
Speed
Queue
Length
Kalman Filtering
Isolated Control
Arterial Coordinated
Control
Regional
Coordinated Control
Reversible Lane
Control
LawEnforcementProcess
Checkpoint
Camera
Red Light
Camera
Human Check
data
Violated
Passed
Double Check
ViolationCheckby
AI
Algorithm
Repository
….
AI Platform
Data Input
Law Enforcement with AI
Feature Extraction Violation Check Double Check
Application
LawEnforcementEcosystem
 Workload: ~300k image capture per Day, but only
process 30k-50k due to workload
 Efficiency: ~70-80% images can not pass the check.
 Algorithm: Deep learning based law enforcement.
Algorithm Repository
 Platform: FusionMind AI Platform
Solution:FusionMind AI
Challenges
 Reduced Workload: Reduce 30k case to 6000 case
for human check.
 New Violation: Modeling time( Reduced from two
month to one day), Process time(Reduced from one
day to 5 minutes)
Business Value
Law Enforcement with Artificial Intelligence
Feature
Extraction
Image Record
For Violation
Not Violated
Truck Restriction
Local Vehicle Only
No License Plate
Improper
Lane Use
……….
Truck Restriction
Local Vehicle Only
No License Plate
Multimodal Learning
IntrafficIn-traffic
In-traffic
AI its all about patterns
Comparison before and after global traffic optimization
Travel Time Delay Wait Time # of Stops
Smart car powered by Mobile phone
SmartCar Smart Transportation
• Autonomous Driving Car:
• Audi Q7 Autonomous Driving Prototype
• Internet of Vehicles
• Based on Huawei OceanConnect
• PSA DS7 Crossback
• PATEO
• 5G LTE-V
• Collaborated with BMW
• V2X: V, I, N, P
• Verified in WuXi, China
Integrate to 10+ algorithm vendors with
20+ AI application.
FusionMind Helps Shenzhen Traffic Police Build Smart Traffic Brain
Cluster resource utilization
improved by 100%
Illegal
occupation
Red light
Signal
indicator
optimization
Video event
Check
License plate
recognition
Roll back the
scrap.
Shenzhen Traffic Brain
... ...
5x
Efficiency improvement
10%
Traffic improvement
Urban road network signal light
Control and optimization
Automatically identify traffic
violations
analysis
algorithms
Video/Image
data source
Analysis
Result
Result
convergence
Algorithm
Evaluation
FusionMind
Smart Traffic
Jorge Sebastiao, CISSP
CTO Ecosystem
DCSS
Twitter.com/4jorge
Linkedin.com/in/sebastiao
How AI is Disrupting
Traffic Management in Smart City

How AI is Disrupting Traffic Management in Smart City

  • 1.
    How AI isDisrupting Traffic Management in Smart City Jorge Sebastiao, CISSP CTO Ecosystem DCSS Twitter.com/4jorge Linkedin.com/in/sebastiao
  • 2.
    Human Society IsEntering the Intelligent Era Data Generated Data Stored Data Processing Actionable Insights Big Data Artificial Intelligence Structured and unstructured data Data is stored in databases and servers Process the data using CPU/GPUs and AI algorithms to detect patterns Predictive signals are generated ICT infrastructures have become the cornerstone of the smart world.
  • 3.
    Big Data +AI  EI Intelligent analysis AI+ Database +Hadoop Big data analysis Hadoop Fusion analysis Hadoop + Database 2013 2015 2017
  • 4.
    Evolution towards EnterpriseIntelligence Theoretical Basis Internal Practice Chip Device Cloud External Enabler (EI services) Noah Ark Lab Distributed Lab Neumann Lab Shannon Lab GTS ISC+ Device BP&IT Heterogeneous computing General AI Essential platforms Domain-specific solutions
  • 5.
    Motivation: Connected Intelligencein Enterprises Collocated storage Device-Cloud collaboration HW-SW synergy Easy to Use Converged decision Full lifecycle support Open & compatibility EI Platform Services
  • 6.
    Easy & Inclusive:Put Enterprise Intelligence Within Reach Feature ranking Feature engineering Model tuning AutoFE & Model Optimization Full-stack AI capability Big data base Intelligent Decision Essential Platforms Generic AI Services Rich services ready to use Full Lifecycle Support Train Evaluate Serving Deploy Upgrade
  • 7.
    Intelligent Computing Platformelements AI Accelerator Module AI Appliance AI Accelerator Card AI Edge Station Heterogeneous Server Intelligent Video Analytics Server GPU FPGANPU
  • 8.
    Pre-integrated solutions forSmart Cities Intelligent Twins  TrafficGo: Traffic optimization, safety, trend analysis  iWater: Sewage treatment, environmental monitoring, river safety City  Production quality:Flexible quality classification, predictive maintenance  Sales forecast: Regional, product, or combined  Production planning: Material forecast Manufacturing  Customer analysis, customer profile, customer forecast, VIP analysis  Abnormal event monitoring: Traffic surge monitoring, illegal behavior (stay/wander) Campus  Inventory management, shipping path planning  Trunk dispatching, smart entrucking  Intelligent declaration: OCR, customs price forecast Logistics  Video recommendation: user profile, video tag, video classification  Content inspection: Pornography, terrorism, violence... Internet  Fleet management: Electric fence, trajectory mining, path planning  Autonomous Driving: MDC, V2X, Scheduling Vehicle  Home Safety: Stranger detection  Family time: Smile moment, pet identification  Elderly care: Fall detection, cry detection, abnormal sounds Home  Pathological analysis: Cervical cancer analysis, lung cancer tissue analysis, pathology report generation  Genetic analysis: Gene sequencing, disease classification and prediction based on genetics Health
  • 9.
    Converged : CrowdIntelligent Decision & Planning city district intersection Hierarchical & Crowd Intelligent Decision lights truck camera RLS MLSGES Converge DLS Smart Transportation Supply Chain Computer Vision
  • 10.
    Urban Transportation Challenges ---Congestion,Violation, Management • Violation Types: Illegal Parking、 Improper Lane Use、Crossing Line(Driving) and Wrong Way(Driving) • Violation Number: ~300k Case per Day in X city • Large Volume of Vehicle:~6 Millions Vehicles in X city. • Congestion:The average speed is ~22km/hour in the rush hour。 Severe Congestion Frequent Traffic Violation Low efficiency Management • Workload: ~300k image capture per Day, but only process 30k-50k due to workload • Scope: Lots of Blind Spot. Non- AI camera.
  • 11.
    Camera Video Signal ControlPolicy AI Based Signal Control Offline Variant Cycle Online Adaptive Video Data Analysis Deep Learning Traffic Data Fusion Huawei AI Platform pedestrian recognition Vehicle recognition Edge Computing Cloud Computing Object Tracking Deep Learning Signal Control: Integrate signal control theory with reinforcement learning to support big area signal control. Edge Computing:Great cost performance with Huawei AI chip Video Analysis:Analyze the pedestrian, Vehicle and Road Structure by deep learning; bird's-eye view video Concise Traffic Data: Fuse the structured data from traditional device with video. Edge Computing Reducing Congestion by Signal Control &Traffic Situation Perception Structured Data from Other Device Lane recognition Tiny Object recognition Image Enhancement Passenger Car Unit Diverging Ratio Vehicle Speed Queue Length Kalman Filtering Isolated Control Arterial Coordinated Control Regional Coordinated Control Reversible Lane Control
  • 12.
    LawEnforcementProcess Checkpoint Camera Red Light Camera Human Check data Violated Passed DoubleCheck ViolationCheckby AI Algorithm Repository …. AI Platform Data Input Law Enforcement with AI Feature Extraction Violation Check Double Check Application LawEnforcementEcosystem  Workload: ~300k image capture per Day, but only process 30k-50k due to workload  Efficiency: ~70-80% images can not pass the check.  Algorithm: Deep learning based law enforcement. Algorithm Repository  Platform: FusionMind AI Platform Solution:FusionMind AI Challenges  Reduced Workload: Reduce 30k case to 6000 case for human check.  New Violation: Modeling time( Reduced from two month to one day), Process time(Reduced from one day to 5 minutes) Business Value Law Enforcement with Artificial Intelligence Feature Extraction Image Record For Violation Not Violated Truck Restriction Local Vehicle Only No License Plate Improper Lane Use ………. Truck Restriction Local Vehicle Only No License Plate
  • 13.
  • 14.
    AI its allabout patterns
  • 15.
    Comparison before andafter global traffic optimization Travel Time Delay Wait Time # of Stops
  • 16.
    Smart car poweredby Mobile phone
  • 17.
    SmartCar Smart Transportation •Autonomous Driving Car: • Audi Q7 Autonomous Driving Prototype • Internet of Vehicles • Based on Huawei OceanConnect • PSA DS7 Crossback • PATEO • 5G LTE-V • Collaborated with BMW • V2X: V, I, N, P • Verified in WuXi, China
  • 18.
    Integrate to 10+algorithm vendors with 20+ AI application. FusionMind Helps Shenzhen Traffic Police Build Smart Traffic Brain Cluster resource utilization improved by 100% Illegal occupation Red light Signal indicator optimization Video event Check License plate recognition Roll back the scrap. Shenzhen Traffic Brain ... ... 5x Efficiency improvement 10% Traffic improvement Urban road network signal light Control and optimization Automatically identify traffic violations analysis algorithms Video/Image data source Analysis Result Result convergence Algorithm Evaluation FusionMind Smart Traffic
  • 19.
    Jorge Sebastiao, CISSP CTOEcosystem DCSS Twitter.com/4jorge Linkedin.com/in/sebastiao How AI is Disrupting Traffic Management in Smart City

Editor's Notes

  • #5 - 呼应前面讲座中提及的华为对EI的理解和定义 介绍华为在AI的相关积累 2002 电信BI业务的数据治理和分析 2007 大数据技术布局和人才积累 2012实验室在基础技术领域的研究 GTS,ISC+等领域的自身实践 最终形成对外赋能的EI服务
  • #6     围绕企业对AI的关注点和诉求,提出EI平台要达到的总体目标          1,Easy & All-inclusive: a) 开箱即用的AI  减低劳动成本、重复劳动; b) 全流程支持的平台  企业大AI需求,如过程智能、自助特征选择、参数调优、易发布     2,Converged: 融合智能决策  企业AI大脑 3,Synergistic: 软硬协同 (异构芯片、系统、网络、GPU集群)  4,Collocated: 为AI提供便利的数据服务  AI数据类型的多样性(tensor, model, graph, image/video…)+ 统一的存储 5,Fully connected: 端边云的协同     6,Open: 生态兼容、开发架构、产权与隐私问题 (数据模型分离)
  • #7 易用性 – 开箱即用 全周期的平台管理,方便从训练到部署和升级的各个过程 基于自动特征提取等技术(电信实践、DL的模型优化) 大量的基础通用AI能力按需使用,REST API方便易用 全栈式的AI能力,满足不同层面的诉求
  • #9 Since AI is still in its early stage Pre-integrated solutions become quit essential to ease AI adoption Given our unique comprehensive capability both on-line and off-line ,cloud and device,hardware and software… The pace of public cloud/hybrid cloud based solution development was so fast that far beyond our expectation It is just one year since Cloud EI was launched, we now have rich pre-integrated solutions available for city, manufacturing, logistic, health…. Etc. With that We feel strong confident to say digital twins as popular term is becoming some thing legacy even still in its infancy Intelligent twins which include all you need on cloud 、edge and device has become the new normal More details will be introduced tomorrow
  • #10 参考智慧城市 – 层次化的、多智能能力的融合决策(知识图、DL, ML, RL) 企业 ->城市 -> 多域融合的诉求 http://3ms.huawei.com/multimedia/docMaintain/mmMaintain.do?method=showMMDetail&f_id=Img201509131942 http://3ms.huawei.com/multimedia/docMaintain/mmMaintain.do?method=showMMDetail&f_id=Img201509290441 http://3ms.huawei.com/multimedia/docMaintain/mmMaintain.do?method=showMMDetail&f_id=Img201405110898 http://3ms.huawei.com/multimedia/docMaintain/mmMaintain.do?method=showMMDetail&f_id=pur200709200004
  • #14 13
  • #15 Please try to add some graphics
  • #17 ISO26262 ASIL-D standard,英伟达这一代产品Drive PX 2都没有达到ISO26262 ASIL-D级别标准 谷歌、苹果、百度、阿里、特斯拉、通用、福特以及英特尔、安波福 352TOPS,并支持L4级别的自动驾驶计算。具体来看,可处理16个摄像头、6个毫米波雷达、16个超声波雷达以及8个激光雷达的数据 四个使能: 联接使能:为汽车提供安全可靠联接,支撑亿级海量连接,同时满足车企业务全球化运营需求; 数据使能:通过对车况和驾驶行为等车辆大数据的采集与分析,在云上实现人和车的数字画像(Digital Twins); 生态使能:通过数据和业务分离结构,帮助车企掌控数字资产,汇聚第三方内容和应用生态,构筑以车企为中心的生态系统; 演进使能:车联网平台与V2X协同发展,从单车智能到车、路协同智能,促进未来智能交通建设。
  • #18 ISO26262 ASIL-D standard,英伟达这一代产品Drive PX 2都没有达到ISO26262 ASIL-D级别标准 谷歌、苹果、百度、阿里、特斯拉、通用、福特以及英特尔、安波福 352TOPS,并支持L4级别的自动驾驶计算。具体来看,可处理16个摄像头、6个毫米波雷达、16个超声波雷达以及8个激光雷达的数据 四个使能: 联接使能:为汽车提供安全可靠联接,支撑亿级海量连接,同时满足车企业务全球化运营需求; 数据使能:通过对车况和驾驶行为等车辆大数据的采集与分析,在云上实现人和车的数字画像(Digital Twins); 生态使能:通过数据和业务分离结构,帮助车企掌控数字资产,汇聚第三方内容和应用生态,构筑以车企为中心的生态系统; 演进使能:车联网平台与V2X协同发展,从单车智能到车、路协同智能,促进未来智能交通建设。
  • #19  深圳是一座拥有2200万人口、GDP总量全国第三的现代化都市,不但是世界上人口密度最高的城市之一,机动车密度同样惊人,全市机动车保有量335万辆,每公里车辆密度为510辆,位居全国第一,人、车、路矛盾更加凸显。深圳交警与华为通过联合创新共建“城市交通大脑”。主要就超带宽交通网络,全市交通流量感知,人工智能辅助执法,提升大数据打击效率,提升市民出行体验五方面展开了深入的合作。 不仅在出行领域,通过公共区域视频监控,15小时找回被拐卖失踪儿童;2017年1月26日 16:00 广东省,深圳市,龙岗区;2017年1月27日 06:00 湖北省,武汉市,武昌火车站。 人工智能辅助执法 现在,深圳交警大数据研判平台实现了对卡口数据运算的秒级响应,基于对车辆外观特征识别的二次识别技术日处理图片能力达到1000万张,对于违章图片的识别达到95%以上。人工智能技术的投入使用,提升10倍的违章图片识别效率,确保了违章图片的闭环处理。 提升大数据打击效率 以前开展一个专项活动需要7天的时间进行数据资源准备、软件开发和数据分析,才能找到合理的数据。现在,深圳交警依托大数据平台及交通分析建模引擎,创建“失驾”、“毒驾”、多次违法等大数据分析模型,30分钟就能形成情报精准推送,开展数据打击专项行动精准查处,定向清除。最近一段时间已经精准查处各类重点违法37055起,查扣假、套874辆,工作效率以往10倍。现在套牌、假牌、报废、多次违法车辆在深圳道路已基本绝迹。 提升市民出行体验 基于交通时空引擎,融合卡口、浮动车等数据,深圳交警已建立全市所有信号交叉口的实时监控系统,制定精准的交通信号管控模式。通过管控大数据,科学设置路口渠化及交通组织创新,道路通行能力力争提高8%左右。 http://www.7its.com/html/2018/dongtai_0821/7323.html 当前道路交通治理的主要矛盾:截至2018年6月底,全国机动车驾驶人数达3.96亿人,机动车保有量达3.19亿辆,驾驶人和机动车快速增多,车辆越来越多与人民群众日益增长的对于美好出行的需求与稀缺的交通资源之间的矛盾,是当前交通管理信息化的主要矛盾,主要表现为当前交通管理信息化系统烟囱林立,各个系统、数据库七国八制,很难实现数据共享和应用协同。面对现状,要实施国家大数据战略,必须要通过多源数据融合,快速提升交通治理现代化水平。 华为的智慧交通云行解决方案,以全栈智能为核心构建“平台+生态”的业务体系,通过感知智能、计算智能、认知智能提供肥沃的黑土地平台,通过业务智能联合合作伙伴让业务应用百花齐放。 感知智能 华为最新X系列智能摄像机的出色功能。当由于速度过快导致车牌模糊时,X系列智能摄像机自动根据车速调整抓拍模式,清晰抓拍车牌,识别车辆信息;当失驾、多次违法未处理、毒驾等违法人员或重点管控人员开车上路时,X系列智能摄像机抓拍高速车内的人脸并识别报。 计算智能 数据越来越多,对算力的要求会越来越高,华为提供了一系列高性能的服务器来支撑交通场景的需求。 华为最新Atlas服务器。在实战场景中,如果要同时处理20000路卡口的视频,传统方式需要至少1000台服务器,而华为Atlas 5500异构计算平台,单台设备实际处理能力在380路以上,处理20000路卡口视频只需要60台,性能至少10倍以上于普通X86服务器。Atlas强大的智能计算能力能够在3秒内完成对1000亿张图片的检索,加速交警对嫌疑车辆的定位,传统服务器至少需要80个机柜,Atals 5500只需要9个机柜,节省机房空间80%以上。 华为的ARM服务器TaiShan和统一存储资源池。华为TaiShan服务器采用了华为自研的CPU,兼容主流操作系统,与华为的大数据平台FusionInsight完美适配,实现真正的自主可控,在北京、天津等地已经投入公安实战应用。同等投入下,TaiShan比传统X86服务器要更加安全,性能提升20%,节省20%空间。面对现在海量的视频数据,传统磁带或蓝光的存储方式无法满足交警实战的需求,华为的分布式智能存储系统单套系统容量可达到100PB,实现计算与存储分离,高密存储、低冗余、低功耗,冷数据随时可用三大特点,单盘14T,单柜11P,是传统磁带库的3倍,节省50%空间,总体成本节省30%。 认知智能 华为的认知智能是一个以“云、大数据、人工智能”为基础,并在生态应用聚合平台提供各种模型服务和算法服务的开放平台。 以深圳交警举例,华为基于开放平台与交警客户进行了深度合作和创新,为解决交通拥堵问题,在华为深圳总部周边进行了红绿灯配时优化,华为基地周边共43个路口,目前已经完成了8个路口。通过AI摄像机感知路口最实时、最准确的路况信息,使用人工智能的方式来认知一个区域整体的交通出行规律,并对该区域信号灯进行整体调优,将原本2~4小时配时一次提升为15分钟配时一次,优化后主干道路平均路口等待时间降低15%。 业务智能 华为与交通领域中的众多合作伙伴建立深度合作关系,华为提供开放平台,伙伴们在上面进行业务创新,一起打造了一个生机勃勃的生态体系,共同建立了5大场景化解决方案: 为了让市民安全出行,我们一起打造了“智能非现场执法”场景化方案 为了让市民有序出行,我们一起打造了“刷脸执法”场景化方案 为了让市民畅通出行,我们一起打造了“TrafficGo”场景化方案 为了进一步提升交警的执法管控能力,我们一起打造了“集成指挥”场景化方案 为了进一步辅助交警提升智能决策能力,我们一起打造了“路况研判”场景化方案