OmniPerception's Computer Vision Technologies Transformational by pkj12584

VIEWS: 5 PAGES: 3

									OmniPerception’s Computer Vision Technologies

Transformational Technology

1. Introduction
      In addition to its world-famous facial biometric solutions, OmniPerception is committed to the on-going
      advancement of its core technologies in a range of other image processing applications. Designed to
      deliver higher levels of performance to customers in mission-critical areas, these solutions fall into two
      broad categories:
          •    OmniPerception’s Magellan™ technology drives a range of products that detect, analyse and
               record the occurrence of logos, brands and other similarly distinctive two dimensional images in
               television footage, video and other image based media;
          •    Under the Gama™ banner, OmniPerception is also advancing specialised derivations of
               Magellan™ technology for mission-critical imaging applications in intelligence, reconnaissance,
               surveillance and video analytics.
      OmniPerception solutions in both areas benefit from distinctive performance advantages rooted in the
      company’s unique core technology. A brief summary of these underlying advantages is set out in this
      paper and the company will be very happy to hear from customers and partners who would like to know
      more (contact details below).



2. The multiple template conundrum

2.1       Apart from the chronic problem of high false alarm rates, the main disadvantage of traditional
approaches to “intelligent video” is that multiple templates are necessary to ensure the detection of objects
appearing at different angles and different real or apparent sizes in the video stream. This requirement
creates a need for very large hardware installations and even then, traditional technology is rapidly
overpowered as soon as the video analytic task becomes at all complex.


2.2       The root of the problem lies in the fact that the appearance of any object or "search target" in a video
stream is greatly dependent on two key factors:


      •   the position of the object in space and
      •    the object's orientation with respect to the viewer (or the camera eye).




                                                                            OmniPerception Ltd.
                                                                            15 Frederick Sanger Road, Surrey Research Park
                                                                            Guildford GU1 2RU, UK
                                                                            Tel: +44 (0)1483 688350 | info@omniperception.com
Each of these two factors can be described according to 3 separate axes or dimensions. That is to say, the
appearance of any one object requires the specification of 3 distinct parameters to describe it fully with
respect to each of these two factors. To describe the object's appearance fully - in respect of both place and
presentation - requires a total of 6 parameters for any one accurate description - or recognition - of the
object.


2.3       One of the most common techniques used for recognizing objects is template matching. However,
template matching only deals with 2 of the 6 parameters involved in the task. It is possible to recognize an
object, but only provided that the object is presented at or very close to the expected presentation. For
multiple presentations, a search system based on template matching needs to use multiple individual search
targets - and search processes. Even with the most advanced modern template matching techniques,
successful object recognition from the use of a single "search target" requires a match to within no more than
15 degrees away from the expected presentation.


2.4       With respect to angle of presentation alone, this means that the system in use needs to cover at least
12 possible angles in each of the 3 ways that the object can rotate. When scale is also taken into account, we
have: 12 different scales, requiring 12-to-the-power-of-4 templates. As can easily be seen, this means that for
fully accurate search coverage 20,736 templates would be needed for each object or search target under
investigation; if conventional template matching techniques were being used.


3. OmniPerception technology cuts the problem down to size

3.1       In contrast, OmniPerception's "Search Space Reduction" (SSR) technology achieves object recognition
much faster and more efficiently. This new approach makes innovative use of distinctive object features to
allow accurate identification with much fewer search templates. Often a single template will suffice, with huge
savings in processing time and hardware cost, as well as significant reductions in false alarm rates.


3.2       OmniPerception’s complementary “Geometric Invariance” (GI) technology allows accurate analysis
even when objects are presented within a huge range of random positioning, sizes and angles of presentation.
This makes solutions based on OmniPerception technology more accurate, more cost-effective, faster and
more reliable, as compared with more traditional video analytic techniques available from competitive
suppliers.


3.3       These factors are particularly important for customers facing the challenges of scene understanding in
intelligence, reconnaissance, surveillance and video analytics. In live tests in the UK and by the company’s
teaming associates in the USA, on TV footage and on footage from surveillance and reconnaissance sources,




                                                                           OmniPerception Ltd.
                                                                           15 Frederick Sanger Road, Surrey Research Park
                                                                           Guildford GU1 2RU, UK
                                                                           Tel: +44 (0)1483 688350 | info@omniperception.com
OmniPerception Gama™ technology has, for instance, been successful in finding, recognising and
differentiating


      •   Cars, large trucks, pick-up trucks, taxis, police cars.
      •   Aeroplanes
      •   Vehicle number plates presented at angles beyond the capability of standard ANPR technology
      •   Other objects of interest


Within these categories, Gama is also capable of finding relatively small wordings or other distinguishing
features on the sides (or other parts) of vehicles, pieces of equipment, cargo items and other objects,
wherever distinctive markings can be found.


3.4       In many cases, the distinguishing features may be characteristic of the objects themselves – e.g. a
vehicle outline or shape of baggage item - and need not be separate markings. Pictographs or “search targets”
need not be pre-selected. They can also be taken from observed footage – for instance the input from a CCTV
camera on one part of a site or premises – and then used to scan footage from elsewhere, both historic and
live or future footage. For example, a vehicle that arouses suspicion can be “tagged” in this way and then
monitored closely via the input from all site cameras (or indeed via cameras in other locations) until action is
taken or the alert is stood down.


3.5       This capability relates to vehicles and also of course to other objects and to people, provided that
some sufficiently distinctive part of their appearance – e.g. their clothing, or what they are carrying - can be
identified. Gama™ can find and identify such search targets even when their distinctive markings are only
partly visible. The technology has an advanced capability in this respect and can be fine tuned to meet specific
customer needs. Where even greater detection reliability is required, the system can also search for a
distinguishing feature in one or several parts, including quite small components provided always that each is
sufficiently distinctive.




                                                                            OmniPerception Ltd.
                                                                            15 Frederick Sanger Road, Surrey Research Park
                                                                            Guildford GU1 2RU, UK
                                                                            Tel: +44 (0)1483 688350 | info@omniperception.com

								
To top