Recent Advances in GPU-Based Intelligent Video Analysis Beijing - - PowerPoint PPT Presentation

GTC'16, San Jose, USA April 5, 2016

Recent Advances in GPU-Based Intelligent Video Analysis

Beijing Vion Technology Inc. Hai Tao, Dr., CEO

Company Introdution

• The Beijing Vion Technology, Inc.(BVT) was founded in 2005,

specialized in computer vision technology. The company currently employs around 200 talented staff members.

• The company has developed products in several markets including

intelligent transportation systems (ITS TS),business intelligence systems, security and urban management systems. These products are deployed broadly across China and BVT is recognized as one of the leading computer vision companies in China.

• Recently, BVT has developed core algorithms for ADAS and UBI

products. San Jose, USA April 5, 2016

A Typical IVA System

San Jose, USA April 5, 2016

Analog Cam Concealed Camera Half Dome Cam IP Cam

IVA Box Network

Switch PC Client Data Manage ment

Mobile App

数据中心

TK1-Based IVA Box - VT-B2081A (CBOX)

San Jose, USA April 5, 2016

TK1-Based IVA Box - VT-B2082A (IBOX)

• 2 TK1 GPU's
• 8 analog video inputs
• 8 audio inputs
• 2 RJ45 1000M ethernet

ports

• built-in GPS
• 1 2.5" hard drive
• Alarm: 4 in, 2 out
• 1 RS232 and 1 RS485
• SPF connection

San Jose, USA April 5, 2016

DEMO: IVA Algorithms Running on CBOX/IBox

San Jose, USA April 5, 2016

• People counting - Shopping mall
• People counting - Summer Palace
• People counting - Bus
• People counting - Subway
• People counting - Cinema
• Urban management - Vending
• Law enforcement - Fighting detection
• Law enforcement - Parking violation

A GPU-Based Smart Camera - VT-E412

San Jose, USA April 5, 2016

A 12MP@30fps GPU smart camera

• The 1st GPU-based smart camera that can reach 12MP @30 fps
• Applications: Face recognition, gender and age recognition,

people counting, LPR, fight and chasing detection, premiter monitoring

• Main features
• 12 MP CMOS sensor @ 30 fps
• FPGA+TK1 ISP pipeline
• Full resolution video analytics
• Sensors: gyroscope, accelerometer, digital compass
• GPS & Wifi/3G/4G supported
• Audio: 1 in, 1 out
• Storage: 2 SD cards
• Alarm: 2 in, 1 out, 1 RS232 and 1 RS485

DEMO: IVA Algorithms Running on VT-E412

San Jose, USA April 5, 2016

• Facial detection
• Gender and age recognition
• Vehicle detection, tracking and recognition

StartNet GPU Cluster

San Jose, USA April 5, 2016

40 GPUs in a 2U Rack Server

• Characteristics of backend IVA
• Handreds of h.264/HEVC video streams from IPCs
• The computation need for decoding and vision algorithm

roughly in 1:1 to 1:10 range

• Using traditional server: not enough matching HW decoding

power, higher FLOPS/Watt, Lower computing density (FLOPS/SpaceUnit)

• Many-GPU solution
• TK1/TX1 balances the HW video decoding/encoding

capabilities with computing power

• Lower FLOPS/Watt
• Much higher computing density

San Jose, USA April 5, 2016

Vehicle barnd/model and color recognition

• Frontal and back view
• f vehicles
• Day/night time
• Average 90% accuracy

for vehicle brand recognition

San Jose, USA April 5, 2016

Driver behavior recognition

• Not fastening

seatbelt

• Using cell phones

San Jose, USA April 5, 2016

Prototype ADAS Algorithm Framework

• Deep neural network based object detection and tracking

(vehicles, pedestrians, bicycles, etc)

• Deep neural network based road scene understanding (traffic

lane detection and following)

• License plate recognition
• ADAS main functions including lane departure warning

(LDW),vehicle distance monitoing warning (HMW), pedestrian collision warning (PCW), and driving behavior analysis (DBA)

San Jose, USA April 5, 2016

Prototype ADAS Algorithm Framework

• Demonstration

San Jose, USA April 5, 2016

Thank you!

San Jose, USA April 5, 2016