New life for rule-based analytics system with Machine Learning
The rapidly-growing global market for mass transit security systems already exceeds $10 billion a year. Increasing instances of violence and rising awareness about mass transit security systems led to its growth.¹
Mass transit security systems gather real-time data on the performance of trains and increase safety with sensors installed near railway tracks. These security systems are used to prevent train collisions and derailments and stop unauthorized access to fenced-off areas.
In this case, a video analytics security system is used by a homeland security company to process visual data from hundreds of video sensors near railroad tracks and extract changes in the images captured. An alert is issued once a suspicious activity is identified by the rule-based system. A challenge arose when the engine routinely mistook clouds, trees and birds for threats. When inundated with false alarms, the control room staff stopped heeding these alerts altogether. Having invested in the physical sensors and video analytics system, this company was looking to leverage this existing investment by improving its performance.
Data Science Question
Can machine learning be used to reduce the number of false alarms the security system issues?
Locations of the sensors vary greatly and examples of truly suspicious activity are scarce. Building a generic model to address this case would have required a huge amount of data to attempt to cover every possible security breach. Such a solution was deemed impossible for lack of data and the difficulty of fine-tuning it for each sensor.
Instead, the approach selected was to train a dedicated ML model for each sensor. This approach allowed for creating a training dataset for each location by simulating events that represent true anomalies.
For this temporal anomaly detection case, Firefly Lab received training datasets representing changes in video images; that data is transformed into a numeric format with a timestamp. Using that data, Firefly Lab automatically created the pipeline and ensemble of several algorithms, including XGBoost, Random Forest and AdaBoost.
The machine learning models automatically created by Firefly Lab are able to reduce the false alarm rate by 90%, well within the goals of the project. More importantly, with this reduction, the security system was made operational again.
Thanks to the automation in Firefly Lab, this homeland security company is able to build a dedicated model per location\sensor without a data science team.