EVITECH crowd monitoring solution Lynx has received field test results from a public transport company, after several months of measures held between late 2013 and March 2014.
EVITECH crowd monitoring solution Lynx has received field test results from a public transport company, after several months of measures held between late 2013 and March 2014. Target of this evaluation was to measure fraud on automatic passengers accesses toward platforms (based on turnstiles), using existing color cameras deployed in stations by the company operating the transport.
The system has been configured to count the difference between tickets validated at the access gates and actual number of people passing through the turnstiles, whether regularly or above them. The interest for such a measure is multiple:
- for statistics,
- so as to optimize days and hours of enforcement,
- to prioritize investment in gates upgrades with additional doors,
- and receive actual reimbursements from public organisms (paying a fee to the transport company on a per passenger basis).
The solution has been tested on typical full days of recorded videos, thus allowing ground truth comparison with manual counting and measure gap between automatic and manual counting. In parallel, the system received the electronic inputs from the turnstiles (one for each paid tickets), to count paying passengers and thus obtaining fraud value.
An individual counting line has been placed on each passage, and in the same time another one in front of the first one to see groups of two and measure the error.
Globally, the picture below provides a synthetic view of the results:
- A 5% mean error was established on an hour per hour basis, given that the system sometimes undercounts (5% below ground truth) and sometime overcounts (5% above ground truth), when compared to manual counting, thus leading to a daily error even lower than 5%.
- Above 100 people were counted frauding on a per hour basis on the two monitored gates
- a fraud rate (nb of invalid entries / nb of valid entries) measured between 20% in the morning up to more than 100% after 9 pm (more fraud entries that regular ones). As the fraud rate is way higher than the measured error on counting, the application is legitimized on this function.
These results are considered very encouraging: "We have noticed a significant improvement with regards to the previous iteration done in may, 2013", indicates the end user, as the previous trials conducted to a precision around 90%, while now it is 95%.
