ACVT Capabilities Show

The Australian Centre for Visual Technologies

Professor Anton van den Hengel, Director ACVTThe University of Adelaidewww.acvt.com.au

Visual Technologies

The production and analysis of visual digital media Primary research areas:

Intelligent video surveillance Enabling thousand-camera surveillance networks

3D modelling from image sets Interactive scene modelling from image sets

Learning from large image sets

Staff

5 Academics 14 Research staff

2 more on the way 10 PhD students 2 Coders 1 Business Manager

Interactive 3D modelling

3D modelling is critical to all sorts of application Special effects, but also mining, architecture, defence, urban

planning, … People are getting more visually sophisticated More 3D data is being generated

More cameras, but also scanners etc The interfaces of modelling programs are usually very

hard to fathom

Modelling

Results

Low polygon-count modeling

Insert your own objects into a game Put your couch into second life Model your house for Google Earth Video editing

Cut and paste between sequencesRemove someone from your home videos

Put your truck into a game

Modelling for animation

Video editing requires models

Modelling architecture

Modeling for virtual environments

Intelligent Video Surveillance

Enabling wide area video surveillance Protection of people and property

Various scales of threat, eg assault to catastrophe Perimeter defence/monitoring Increased detection of theft, reduced false alarms

More effective use of police and other security resources Road traffic analysis Public liability

Archived footage provides defence against spurious claims

Camera surveillance is ubiquitous London citizens viewed

>300 times per day >1M cameras in London

Cameras monitor streets, shops, airports, etc

What happens to the video? Most video is never seen Little observed in real time Forensic value only

Understanding the relationships between cameras fundamentally changes what you can do with video Live analysis Forensically


Live human monitoring needs to focus on key events Need systems need to detect such events Scalable solutions Allows more effective responses

Forensic analysis needs to be faster and smarter Need semi-automated data mining

What we are creating is a virtual director, who gives the operator the most important information available at any moment



Camera topology Geometric relationship

Activity topology Characterise movement of

targets between video fields of view

Tells you where to look when a target leaves the field of view of the camera you are currently watching

Outputs from activity topology analysis: Knowledge of which regions are linked Given an object’s arrival in a region, the relative

likelihood that it came from any of the connected regions

Statistics re time taken to transit between regions

Information re typical appearance changes between regions (scale, brightness, etc)



Abstract over individual cameras

Present only the most important information at any time

Effective exploitation of automatic and human analysis

Natural Augmented Reality

Augmenting reality requires accessible models of real objectsCanned models are not

practicalNeed to create them in-

situ

Natural Augmented Reality

In-situ modelling

Live effects

Using AR to put special effects into video, as it’s capturedCopy and paste objectsDelete objectsAdd synthetic geometry

Minimal interaction AR modelling

Augmented Reality

The interaction between real and virtual objects is a key problem in AR

Modelling can solve this problem

Augmented Reality

Modelling plants

Working with The Australian Centre for Plant Functional Genomics, and the Plant Accelerator Hundreds of plants with

differing genes grown on a conveyor belt

Interested in how environment affects yield

Modelling plants

Developing methods for consistently labelling parts of plantsOver timeOver different plants

3D models are a by-product

Modelling plants

Plants Have complicated shapesOverlapMoveGrowDieHave a grammar

Modelling plants

Approach3D volumetric modelsMultiple hypothesis

labellingReconcile models over

image pairs

Simulation

Serious Games Working with

BHP BillitonUNSWTAFE…

Search in large image sets

Developing methods capable of extracting information from large image databasesGoogle earthStreetviewSurveillance video


Databases of > 106

images or video Image or text queries Results in seconds

Aerial surveillance

Unmanned Aerial Video

3D modelling from video

Unmanned Aerial Video Generate a topographic

map based on video taken from an unmanned aerial vehicle

Relies on video only No GPS, lidar, inertial …

Terrain reconstruction from UAVideo

ACVT Experience

Development of iOmniscient IPNow the highest selling intelligent video surveillance

product world-wide Research contracts with DSTO, Canon (CiSRA) Collaborations with Tenix, Coastwatch … Development of networked camera surveillance

platformPatent pending

ACVT Capabilities Show

Technology

Transcript of ACVT Capabilities Show