Page 68 - Security Today, May/June 2019

Everything in Between Edge computing: the evolution of video content analytics
BTy Tom Edlund
oday, artificial intelligence and machine learning are backing a wide range of technologies and applications, powering diverse solutions to a broad set of challenges. For video analytics, deep learning has accelerated the technology’s evolution, particularly when it comes to
accurate detection. AI-backed video analytics enable object extrac- tion, recognition, classification, and indexing—activities that can advance various business and security applications by making video searchable, quantifiable and actionable.
This article will review the factors driving the adoption of deep learning-based video content analytics, the creation of more sophisti- cated cameras and higher resolution video and, ultimately, the result- ing need to identify efficient data processing and computing solutions to support these changes.
Video-Based Alerting:
From Computer Vision to the AI Age
When video analytics first emerged, products were primary designed as alerting solutions. Through triggering calls to action, these early solutions were attempting to eliminate the need for active human vid- eo monitoring. However, these computer vision-based solutions did not fully achieve the aim of removing human involvement in video surveillance and oversight: For one, these video alerting technologies tended to produce false positives and inaccurate matches for video search criteria.
An alternative approach adopted by other solutions was to main- tain and maximize human involvement in the video surveillance process: These interactive video solutions didn’t focus on entirely re- moving human operators from surveillance monitoring but strived instead to accelerate video review for users and make it easier to understand whole scenes captured by video. While alert-based video monitoring yielded imprecise results, solutions that streamlined us- ers’ comprehension of entire video scenes enabled operators to over- come video-based alerting limitations and quickly identify critical information in captured video.
The Renaissance of Alerting:
New Innovations and New Challenges
The introduction of deep learning-backed video analytics revolution- ized the video content analytics industry, driving more accurate detec- tion capabilities and precise alerting. The demand for higher quality video analytics, among other considerations, catalyzed the develop- ment of more sophisticated cameras, as well as end user adoption of more cameras to optimize real-time alerting. These developments have enabled capabilities such as people counting and face recog- nition-based alerting: Higher resolution video makes it possible to more accurately distinguish between people in crowds and capture individual faces, which could then be analyzed by state-of-the-art analytics to trigger real-time alerts when certain conditions are met. Furthermore, beyond alerting, deep learning-driven solutions make it possible to leverage the valuable and powerful video metadata to drive deeper insight in other ways, such as business intelligence and trend visualizations.
While driving the deployment of real-time, deep learning-based alerting solutions, the proliferation of cameras with higher resolution video also drove up the total cost of ownership for video surveillance integrations. Specifically, these conditions entail higher processing
34
MAY/JUNE 2019 | SECURITY TODAY
demands and hardware requirements. For real-time video analytic solutions such as face recognition alerting, better accuracy drives up operating costs—a new challenge that must be overcome.
Lowering the Increasing Cost of Computing
Current video analytics research and development is focused on low- ering the cost of processing. Whereas today’s deep learning driven video content analytics are mostly based on GPU computing, look- ing forward, solution innovators must consider continual improve- ments to camera technology, increasing availability and volumes of high-resolution video, and powerful, deep-learning-driven video ana- lytics, and determine which processing model could best keep costs down: edge or centralized computing?
Because there are advantages to both options, it’s important to understand what makes edge processing and centralized computing respectively effective. Today’s leading solutions rely on centralized computing for several compelling reasons.
Flexible resource allocation. Today, organizations rely on large video surveillance installations with multiple cameras. At differ- ent times, each camera will have varying levels of activity, and by distributing processing with centralized computing, lagging can be prevented. Centralized computing is flexible, enabling the sharing of processing resources between cameras so that unusually high activity can be processed without slowing down computation across cameras. Statistically, relying on more video streams increases the likelihood of maintaining a steady state.
By contrast, edge computing is rigid, requiring pre-defining com- putation resources and scenario specifications, such as activity and resolution, which are dynamic conditions. When working with edge de- vices, users or developers must decide up front whether to provision for normal situations—in which case there is a risk of lagging and missed alerts during high activity scenarios—or for extreme situations—driv- ing up costs, because resources are often idle and do not require the allocation of high processing resources.
When one camera is driving up activity and time is of the essence, overloading processing requirements could cause alerts to be delayed or missed when they matter most. By distributing the computing, lags in alerting can be prevented, timely processing is ensured, and lower processing costs are maintained. The ability to flexibly distribute pro- cessing with centralized computing is more beneficial to deployments with more cameras.
Broader coverage of analytic capabilities. On-camera analytics require pre-configuring the specific analysis activities for edge pro- cessing. Edge devices are typically designated for dedicated purposes, and the range of analytic activities that can be completed per device is limited. Because of device memory constraints, at the onset of de- ployment, the user will need to manually configure the relevant ana- lytics based on the camera location. If the camera points to an area where faces can be viewed in high resolution, the device will likely be dedicated for face recognition, but not for license plate recognition (LPR) for fear of overbearing the processing load.
With centralized processing, there is no need for manual calibra- tion. There is sufficient memory to share different Deep Neural Net- works between video streams and cameras, so that when a person of interest on a pre-defined watchlist passes a dedicated LPR cam- era—capturing a high-resolution image – a call to action can still be
ARTIFICIAL INTELLIGENCE