Sensitivity of a Video Detection System and the Impact on Availability vs Safety

Automated and unattended transit systems form integral parts of many airports (Airport People Movers Systems, APM) but also urban transit systems (subways). Most of such systems have onboard CCTV technology that can be used to determine the status onboard the vehicle, i.e. whether it is empty or not. Passengers, luggage or any other kinds of objects shall not remain onboard when the vehicles are removed from service, e.g. send to the depot or to the workshop or parked alongside the track. A video-based detection system has been developed to automatically detect passengers or objects left behind. Such as with all detection systems, thorough calibration of the sensors, i.e. cameras in our case, is essential to minimize false alarms but also achieve a very high detection rate. This paper presents an empiric and heuristic approach to analyze the detection performance and choose suitable parameter settings to achieve both a low number of false alarms and a low number of wrong side failures to detect every object which is present. because they impact the systems' availability and to provide no wrong side failures to detect every present object). The interdependency between operational availability which is impacted by the number of false alarms and the detection accuracy not to fail to the wrong side is presented based on numerous (hundreds) of real-life tests of an APM system. The results provide insights into the challenging calibration of a detection system and can help understand the contradictory requirement to satisfy availability and safety. The detection performance of this system was analyzed with collected video footage of different test cases (people, groups of people, small and large objects). A specific system configuration was identified which can provide both an accurate detection (higher 99%) and very few false alarms (less than 0.5%). The interdependency between system availability and safety is revealed, i.e. the higher the detection accuracy, the more false alarms appear and vice versa. A “phenomenon” that some practitioners may not be well aware of, even due to the fact, that real-world systems cannot be used as lab-installations to modify parameter settings and reveal this interrelation. In the future, more systematic search strategies for optimum parameters could be implemented and can use the heuristically found parameters as initial solution. Numerical search algorithms can then help solve this tradeoff problem in an optimal way.