Vigilant Real-time storage and intelligent retrieval of visual surveillance data Dr Graeme A. Jones

1
Vigilant Real-time storage and intelligent
retrieval of visual surveillance dataDr Graeme
A. Jones
V
2
Vigilant aims

• to design a database that provides real-time
  efficient storage of events occurring within a
  monitored scene.
• to enable untrained security operators to
  generate human-centric queries for video data
  i.e. queries based on content.

3
System constraints

• Handle video stream in real-time
• Compress the terabyte of digital video generated
  per camera per day onto an swappable one gigabyte
  disk
• Maximum of one high spec PC per camera
• Intuitive button-press query builder

4
System architecture
5
System architecture
6
Fast Robust Event Detection (and tracking)
Algorithm

• Model-based approach exploiting expected
  projected object size, with shadow/reflection

7
(No Transcript)
8
FREDA- Low-level

• Blob detection from pixel comparison with mean
  and variance image

• Temporal updating of reference images with soft
  gating

9
FREDA - High Level

• Region detection from connected components

• Unsupported regions generate new objects

• Current objects validated from supporting regions

10
Example Results
11
Information Balance Sheet

• Input data
• 800x600 x (10.50.5) x 25 x (60x60x24)
• ? 2Tbytes/day
• Output knowledge
• Periodic background DCT updates
• Subimage sequences (pixels and contour) per
  temporal event

12
The Intelligent Camera

• Boundary between PC and intelligent camera
  depends on issues of frame rate, bandwidth and
  computational resources

Pixel Comparsion Connected Components Hypothe
sis Generation Object Validation
Characterisation
13
System architecture
14
Object Classification

• Object event may be classified into Person,
  Vehicle, Large Vehicle classes based on history
  of depth-compensated dimensions and speed.

15
Object Classification
16
Colour Annotation

• Dominant colour(s) of object extracted as modes
  from colour histogram generated from pixels of
  temporal event.

HUE
Munsell Space. Semantic classification in HSV
colour space
Value
17
3D Trajectory

• Ground plane calibration (learnt) enables 3D
  speed to be computed, and hence velocity
  behaviours derived e.g. car maneouvring, person
  running, etc.
• Trajectory commentary derived from areas of
  interest previously assigned labels by operator
  e.g. gate, bikeshed, disabled parking.

18
083456 Vehicle 3434 enters gate 083503
Vehicle 3434 enters F.Lane 083512 Vehicle 3434
enters carpark 083531 Vehicle 3434 enters
zone3 083605 Vehicle 3434 stops in
zone3 155523 Vehicle 3434 leaves zone3
19
Behaviour analysis

• Hidden Markov Models based on states derived from
  clusters of positions along training sets of
  trajectories used to classify object and its
  behaviour

• Car Entering
• Car Leaving
• Person Entering
• Person Leaving

20
System architecture
21
Conclusions

• Its easy to derive atomic units of useful
  user-oriented knowledge
• Requirement for human-oriented query
  specification tools employing fuzzy matching
• Plug n play characteristics e.g. camera
  calibration
• Distribution of computer intelligence

22
Contact Details

• Email g.jones_at_kingston.ac.uk
• Web www.kingston.ac.uk/dirc/