Page 10 - Security Today, October 2020

Driving New Opportunities
The advantages of greater processing power at the edge and greater scalability in the cloud
By Robert Muehlbauer
When network video cameras first came on the market, they were chiefly bare bones video stream- ing devices. Most of the intelligence and processing for the system was housed in the core server farm of the video manage- ment system. Within a few years, however, companies were manufacturing cameras with enough CPU power to perform simple analytics at the edge. As comput- ing power continued to increase, so did the opportunity for companies to embed ever more sophisticated analytics in-camera.
There were several benefits that made edge-based analytics appealing:
Lower bandwidth consumption. Instead of streaming every frame of raw video to the server for analysis, the camera could pre-process the images and just send the event footage.
Lower storage requirements. With only content-rich video being sent to the server, there would be less footage to archive in the
storage array.
Lower operating costs. Processing video
in-camera was less expensive than monop- olizing CPU cycles on the server.
The earliest algorithms brought into the camera were based on pixel changes in the field of view. When the changes reached a certain threshold, the analytics would con- clude that motion had been detected and would send the video to the server. Build- ingonthatpixelthresholdconcept,other in-camera analytics like camera tampering and crossline detection soon followed.
HOW MACHINE LEARNING IS IMPACTING ADVANCES IN VIDEO
Fast forward to 2020. Manufacturers are building cameras embedded with deep learning processing units (DLPUs), en- abling software developers to integrate arti- ficial intelligence (AI) into their video ana- lytics algorithms. This has raised new hopes that machine learning and deep learning will be the silver bullet that the security in- dustry has long been promising. Given the
variabilities of surveillance environments, however, fulfilling that promise still has a way to go. That is because machine learn- ing can consume an enormous amount of resources before a consistently accurate re- sult can be achieved.
We’ll use the example of facial recogni- tion. If you wanted to create the applica- tion using AI, you would need an iterative process to train the program to classify an imageasaface.Thatwouldmeancollect- ing and labeling thousands of images of facing and feeding them into the program, then testing the application after each cycle of input until you determined that the pro- gram had learned “enough” about what characteristics comprise a face. At that point, the trained model would become the finished program. But after that, the AI wouldn’t learn anything new.
Now consider the challenges of facial recognition from a surveillance camera per- spective. Not only do you have to train the program to recognize a full-frontal image, but images captured from multiple angles,
10
1020 | SECURITY TODAY
VIDEO ANALYTICS
Gorodenkoff/Shutterstock.com