Bicycle Theft Detection Motivation and Prototype Dima Damen and David Hogg Computer Vision Group Jan 22 nd, 2008 1
Facts l 500,000 Bicycles stolen annually in the UK l 21,236 bicycles stolen in London (2006/7). l 5% of the stolen bicycles returned to their owners. (2005) l Highest rate of bicycle thefts in: the Netherlands, Sweden, Japan, Canada, New Zealand, England, Finland and South Africa Jan 22 nd, 2008 2/30
Facts Source: EUICS report, The Burden of Crime in the EU, A Comparative Analysis of the European Survey of Crime and Safety (EU ICS) 2005 Jan 22 nd, 2008 3/30
From the news l 7/6/2007: York (290 bicycle thefts during May 2007) city sets up CCTV cameras over bicycle racks. l 22/6/2007: Oxford (1800 bicycle thefts during the last year) city sets up CCTV cameras over bicycle racks. Jan 22 nd, 2008 4/30
From the news l 23/5/2007 Catching Daniel Westrop have been stealing commuters' cycles, often two a day, for the past three years!! Jan 22 nd, 2008 5/30
What can Computer Vision offer? l Tracking l Object Detection l Change Detection l Colour analysis We still Can t do Behaviour Analysis Night tracking Jan 22 nd, 2008 6/30
Computer Vision at University of Leeds l Over 15 years Jan 22 nd, 2008 7/30
Associating Drop-offs with Pick-ups Jan 22 nd, 2008 8/30
The approach Identifying Sequences of Events Drop off Pick up Jan 22 nd, 2008 9
Associating Drop-offs with Pick-ups 1. Tracking People 2. Detecting Bicycles 3. Deciding on drop-off and pick-up actions. 4. Comparing colour information 5. Raising warnings Jan 22 nd, 2008 10/30
Associating Drop-offs with Pick-ups 1. Tracking People 2. Detecting Bicycles 3. Deciding on drop-off and pick-up actions. 4. Comparing colour information 5. Raising warnings Jan 22 nd, 2008 11/30
Associating Drop-offs with Pick-ups 1. Tracking People 2. Detecting Bicycles 3. Deciding on drop-off and pick-up actions. 4. Comparing colour information 5. Raising warnings Jan 22 nd, 2008 12/30
Associating Drop-offs with Pick-ups 1. Tracking People 2. Detecting Bicycles 3. Deciding on drop-off and pick-up actions. 4. Comparing colour information 5. Raising warnings Jan 22 nd, 2008 13/30
Associating Drop-offs with Pick-ups 1. Tracking People 2. Detecting Bicycles 3. Deciding on drop-off and pick-up actions. 4. Comparing colour information 5. Raising warnings Jan 22 nd, 2008 14/30
Resolving Ambiguities Person 1 Bike 2 1 3 4 5 6 2 Jan 22 nd, 2008 15/30
Resolving Ambiguities l Deferred Logic l MHT l MCMCDA Jan 22 nd, 2008 16/30
Experiments & Results l 3 experiments l 1 hour (45 events) l 50 minutes (22 events) l Full day (9 hours and 30mins) (40 events) Jan 22 nd, 2008 17/30
Application to bicycle theft detection Predicted Actual Thief Non-Thief Thief 11 2 Non-Thief 17 183 Jan 22 nd, 2008 18/30
Strengths & Weaknesses Strengths: l Decrease in required monitoring time. Weaknesses Recorded time: 11 hours and 30 minutes Warning time: 13 minutes Jan 22 nd, 2008 19/30
Strengths & Weaknesses Strengths: l Decrease in required monitoring time. l Raises warning, no action taken. Weaknesses Jan 22 nd, 2008 20/30
Strengths & Weaknesses Strengths: l Decrease in required monitoring time. l Raises warning, no action taken. Weaknesses l Person changing clothing. Jan 22 nd, 2008 21/30
Strengths & Weaknesses Strengths: l Decrease in required monitoring time. l Raises warning, no action taken. Weaknesses l Person changing clothing. l Does not detect suspicious behaviour. Jan 22 nd, 2008 22/30
Strengths & Weaknesses Strengths: l Decrease in required monitoring time. l Raises warning, no action taken. Weaknesses l Person changing clothing. l Does not detect suspicious behaviour. l Warning is raised after the bicycle is removed. Jan 22 nd, 2008 23/30
System s Failure Cases 1. The thief wears the same clothing as the owner. Jan 22 nd, 2008 24/30
System s Failure Cases 2. The thief drops another bicycle and picks a better one at the same time. 3. The tracker loses track of people as they pause 4. Theft cases of parts of the bicycle Jan 22 nd, 2008 25/30
We actually caught thieves!! Jan 22 nd, 2008 26/30
Conclusion l 77% theft detection rate. l 8.5% false negative rate. l 1.9% of required monitoring time. Jan 22 nd, 2008 27/30
Publications Damen, Dima and Hogg, David (Sept 2007). Associating People Dropping off and Picking up Objects. British Machine Vision Conference (BMVC 07). Damen, Dima and Hogg, David (July 2007). Bicycle Theft Detection. International Crime Science Conference. (CS2 07) http://www.comp.leeds.ac.uk/dima Jan 22 nd, 2008 28