Computer Vision for Automatic Traffic Monitoring
March 2018
Group of Prof. Yehoshua Y. Zeevi
2
Agenda
• Context
• Goals / Vision
• Technologies / Infrastructure
• Proposed solution
• Challenges
• Results
• Conclusions and Recommendations
• Further work
Context - Smart City
Smart City
Smart Infrastructure
Traffic Control
Safety Efficiency Modularity Security
Context - Smart City
5
Focus – Pedestrian Monitoring
Demo
• Traffic monitoring / Junction overview
• Vehicle monitoring
• Pedestrian monitoring
• Pedestrian safety
• …
7
Goals
8
Possible Technologies
Sensors: motion sensors, magnetic sensors, cameras, BT, LIDARs…
Connectivity: wired / wireless, PTP, IoT, cloud services, …
Computing: low cost/energy computers, PCs, GPUs, cloud computing, …
Computer Vision: simple algorithms, advanced algorithms, deep learning
Haifa Municipal Traffic Control Center Cameras
Raspberry Pi board with HD camera (~100$)
ODROID (~200$)
NVIDIA Jetson GPU (~600$)
INTEL NUC (100 – 500 $)
9
Infra Structure
Simple• Spatial Operators• Time OperatorsAdvanced• Machine Learning
• Object Detection• Object Recognition• Object Tracking• Scene Understanding
10
Computer Vision Tools
11
Control center
Proposed Solution
• IoT (Internet of Things) Methodology
להחליף לרמזור הולכי רגל
12
Golda junction – prior data
Lane recognition
Vehicle counting per lane
Starting PositionVehicle traffic monitoring
13
Chorev junction
Vehicle traffic monitoring
Main FocusPedestrian Monitoring
• Pedestrian counting
• Average waiting time
• (abnormal behavior detection)
Spatial & Time Operators
Person No Person
Machine learning
* Videos from Haifa Traffic Control Center
Pedestrian counting
Pedestrian monitoring
* Videos from Haifa Traffic Control Center
Pedestrian monitoring algorithm
* Videos from Haifa Traffic Control Center
Pedestrian Counting
• New video data• Taken with Raspberry Pi camera
at Haifa Ziv junction
Some statistics
Person People Empty
TPR(%) PPV(%) ACC(%) TPR(%) PPV(%) ACC(%) TPR(%) PPV(%) ACC(%)
Nesher1b 83 85 96 87 94 97 98 97 97
Nesher4a 98 92 99 100 94 99 98 99 98
Ziv1a 98 93 97 98 95 98 95 97 96
TPR(%) PPV(%) ACC(%)
Average Std Average Std Average Std
Person 93 9 90 4 97 2
People 95 7 94 1 98 1
Empty 97 2 98 1 97 1
Challenges
• Sensitivity to viewing angles
20
Control center
IoT Cloud Services
• IoT (Internet of Things) Methodology
Chorev Junction Mapping 1
cam01
cam02
cam03
cam04
cam05
cam06
cam02 cam04
cam06
cam07
cam08
cam09
cam10
cam07
cam08cam10
cam09
Chorev Junction Mapping 2
Junction Diagram
M1
M2
M3M4
M5
M6
M7
M8
E1
E2
E3
E4
E5
E6
E7
E8
E9
Streamed Information
Control center
No images
No infringement of privacy
Results
26
Control center
Into the Future
• IoT (Internet of Things) Methodology
Smart Traffic Lights Standardization
Use of IoT
Conclusions
• Standardization: Traffic lights with built-in cameras
• Distributed computation• Low cost smart sensors• Connectivity through cloud services• Availability to nearby vehicles• Identification of behaviour with reference to age
Current Activity
Tel Aviv
Traffic Light Control - Habima
Disabled Pedestrian Monitoring
The Team
Principal Investigator:
Prof. Y. Zeevi
Researchers :
Dr. Eli Appleboim, Dr. Israel Berger, Johanan Erez, Dr. Rami Cohen, Roy Miterani, Aviad Levis
Vision and Image Sciences Lab :
Daniel Yagodin, Ina Talmon, Aviel Avraham, Alon David,
Zvi Lederer, Ben Ajami
Networked Software Systems Lab:
Roy Miterani, Hovav Gazit
~15 students:
Tom Shitrit, Yan Yampolsky, Lior Haimovich, Roi Sinoff, Tali Srebo, Ohad Spitzer, Alon Mamistavlov, Yonatan Shlain, Yonatan Shahor, Amit Enoch, Amit Gacket, Emanual Alkobi, Nofar Mann, Alon Zabatani, Morag Tohamy
Technion Transportation Institute
Dr. Ayellet Gal-Zur, Oded Komar, Iliah Finkelberg
Haifa Traffic Control Center
Anat Gilad, Sharona Cohen
30
Field Tests
31
Field Tests