Video analytics competitive tests in a metro environment

AN2V meeting of June 25th, 2019 : experience feedback about video analytics solutions when used in public transports

New metroEVITECH was presenting today at the AN2V* meeting about "Innovation in security solutions technologies" an experience feedback concerning a comparative test of crowd video analytics solutions. This test was applied on an international new large metro site. Practically, four international suppliers were compared, for 5 analytic functions tested : panic move, people grouping, bottlenecks due to standing crowd, reverse walk, and immobility (laid objects, staying people). For around 10 test days, the transport operator simulated events with plastrons, and also captured video of such events that could happen "from real life". Several technologies including Algorithmics and/or Deep Learning were tested one against the other on the same images.

LYNX EVITECH software has demonstreted a capacity to detect 100% of the searched events with between 0 and 4 false alarms per day and camera (2 to 3 in average) on each scenario. If this performance rate was not reached over each camera, it was due to the difficult test conditions (limitation of the access to images to one single point abroad, no possible export of images to our R&D team, during the configuration phase) as well as some difficult scenes (running carpet, shining marble) that have limited the detection rates. This has also drawn improvement opportunities, for improving LYNX.

In the comparative analysis, beyond these results, LYNX solution also presented very strong points such as the possibility to address all the project's scenarios with one single software solution, the capacity to justify the nature of each alarm (no confusion) as well as the capacity to to locate an alarm origin in the image with a red frame. This last function is very important for the surveillance agent, who has the doubt raising responsibility : he immediatly knows where to look at. In this domain, general purpose anomaly detection solutions still have progresses to accomplish before well tracing the origins of their alarms (nature and location).

Image illustrating the type of testing site.

* AN2V : French National Association of Video surveyed Cities.