Field trial - AI video analysis for perimeter protection

2021.04.09

From January to March 2021, Securify evaluated U-alarm, a software for perimeter protection developed by Ultinous. 

U-alarm is based on one of Ultinous’s in-house developed datasets and AI video analysis and is a further development of the previous software version named U-vap. The data set for U-alarm has been trained with just over 8 million images to classify people in outdoor environments. 

Multi-level system: 

• Step 1. Radar for detection and tracking of movements
  

• Step 2. PTZ camera for image verification of detected movements
  

• Step 3. AI video analysis for object classification

The field test began in early January -21 when Securify gained access to a BETA version of the U-alarm. The objective was to test U-alarm in as realistic and challenging circumstances as possible. Not the standard setup used by many to showcase their products, using their back yard or other close perimeters as a scene.
Instead, we decided to perform video analysis on a video stream from a moving PTZ camera, controlled by a radar system. No presets were used, the PTZ camera was dynamically controlled by the radar system to follow detections qualified by the radar system. This meant that the conditions for U-alarm were the most difficult imaginable, a significant difference from analyzing video from fixed surveillance cameras with only a 40-60 meters range. 

The test system was installed in a demanding environment and an extensive area that would be monitored with a lot of activity, rail traffic, service staff, deer, rabbits, buildings, fences, trees, a lot of vegetation, and uneven lighting. The radar system's detection area covered approx. 7 hectares while the alarm zones, mainly covered areas within the fence line, approx. 3 hectares.

During a randomly selected measurement period at the end of January, the radar system qualified a total of 699 detections during 72 hours. So, during this period, the radar system activated the PTZ camera to follow qualified tracks, a total of 699 times. This in turn meant that the Ultinous U-alarm analyzed video from 699 unique events. Of these 699 events, U-alarms classified 65 events as human. Of these 65 events, we registered 5 fails. 

All classifications took place during night hours and with relatively poor lighting. Also, worth noticing, all classifications were done at a distance from 150 to 250 meters from the position of the camera. Some classifications that stood out from the crowd:

• A train driver inside an arriving train at night was successfully classified

• Train technician who, for some reason, decided to ride at the front of a moving train
  

• A person outside the fence line at a distance of 390 meters (this one during daytime)
  
  

Briefly about the system:

Video analysis: Ultinous, U-alarm ver 1.0.1

Camera: Axis, Q6215-E (built-in IR with up to 300 meters range)

Radar sensors

Radar server

VMS

About U-alarm: 

U-alarm belongs to the software category 'AI video analysis' which with the help of machine learning classify different objects in real-time. U-alarm is in this version trained with more than 8 million images to recognize and classify humans, walking, crawling, well in all kinds of postures. Unlike previous generations of video analysis (more like advanced motion detection), U-alarm does not require any perspective settings and other time-consuming configurations, nor does it require any adjustments related to seasonal changes. Bottom line is, U-alarm is both scalable and reliable and should be a good fit for clients with high-level expectations regarding false alarm rate (FAR) and probability of detection (POD). 

This technology is relatively processor-intensive and for this reason, the analysis takes place on a dedicated server with an Nvidia GPU with 256 CUDA processors. An advantage of server-based analysis is that it is camera-independent and analysis can take place regardless of which camera manufacturer and camera models the customer prefers. The analysis takes place on a separate video stream with lower resolution and frame rate; 800 x 450 px and 2 fps. In the current version, U-alarm has support for the analysis of day & night cameras, both fixed and PTZ cameras. Ultinous plans to introduce support also for the analysis of thermal imaging cameras.