Page 23 - Security Today, March 2020

“AI-based cameras aren’t inherently doing anything a human can’t do, but they are able to make associations and recognize certain objects exponentially faster.”
By Ray Cooke
er to identify and classify objects.
Once the algorithms have been pre-trained to identify cer-
tain objects and characteristics, they do not have the capability to “learn new things” by themselves. Additional capability is de- ployed by repeating the above process and deploying new firm- ware to the camera and its deep learning accelerator resources. This allows for a lighter approach to AI hardware resources on the camera, and still allows strong capability to deploy to the edge which can get stronger as it evolves.
A deep learning algorithm doesn’t see video in the same way we do, but it can look for familiar shapes and patterns that it has been trained to recognize. It can’t think for itself or make deci- sions that it hasn’t specifically been programmed to make. AI- based cameras aren’t inherently doing anything a human can’t do, but they are able to make associations and recognize certain objects exponentially faster.
Because an AI model doesn’t understand the context of a situ- ation in the same way that people do for many use cases and com- plex situations, it will be quite some time before the technology can reliably decide and take action autonomously. It can, how- ever, reliably show us events that are likely to be worth human attention, and when appropriate, feed the events to other systems to add more value.
Applications for Security
For security applications, there are two main areas where camera- based AI will significantly improve and enhance security operations:
1. Identifying alert-worthy events
2. Enhanced Forensic Search Post-event
Identifying alert-worthy events. Traditional video analytics that send alerts or tag motion events such as line crossing, loiter- ing and object left behind are prone to errors and false positives from wind, rain, or people standing in front of the object in ques- tion. These previous generation video analytics only see “motion blobs” as opposed to objects with properties that allow them to be classified.
By using AI to detect and identify specific object types like people or vehicles, we can greatly reduce false alarms, while ig- noring things like wind, rain, shadows and an errant plastic bag floating by. AI enables an entire new class of analytics, with more sophisticated logic and customization for precisely what an end user requires.
AI can also help us count objects like people or cars more pre- cisely . This includes the ability to count objects accurately even when they partly “occlude” or pass in front of each other. This is key since it allows use cases like people counting from more sen- sible camera view angles. This is far beyond today’s video analyt- ics, which require a top-down view to avoid occlusion, and which gives a less useful camera view when you want to see faces.
Forensic search post-event. Beyond event triggers and object classification, it is important to realize how much descriptive metadata an AI-based camera can capture with each frame. And
because the metadata is small, it adds very little to the overall bandwidth and storage requirements.
Several defining characteristics of a detected object can be captured such as the color of a person’s shirt and pants, length of garment, hat or no hat, glasses or not, handbag or not, and approximate age and gender. The impact on forensic search is profound. Imagine the time it takes to search through 10 hours of video looking for a man with a blue shirt and shorts. With the embedded metadata provided by an AI-based camera, the search yields results within seconds.
AI technology can even extract clues to behaviors like falling down or fighting by using human skeletal characteristics to classify how people are positioned. Being able to search through this em- bedded metadata on a VMS will require a plugin or API that can read the data. This capability will be available in Hanwha’s Wisenet WAVE VMS and plugins are also being developed for VMS sys- tems such as Genetec’s Security Center that make integration easy.
Applications for Business Operations – Going Beyond Security
With all of this collected data, there’s never been a better time to start thinking about video cameras in a broader sense. AI-based cameras will become important data collection sensors which go far beyond simply capturing video.
They can identify and count objects, display heat maps and ensure process and operational compliance. As a result, they will become an invaluable tool for business operations. Depending on the business, the value proposition for such data can be a game- changer worth many times the cost of the system. Cameras are already well accepted and commonplace, so the opportunities for them to evolve into unobtrusive, important data gathering tools for business and operations intelligence will only continue to grow.
Once seen as merely protecting the bottom line from loss, AI cameras can truly be seen as an enabling technology for revenue generation. They can be tools for operations and process measur- ing and metrics.
Bringing AI Data back from the Edge
With AI on the edge, the valuable events and other metadata gen- erated by cameras will need to be gathered from many endpoints and the data aggregated together, usually in server or cloud-based systems, to enable clear visualization of the trends and anomalies identified. This can be presented via a suite of basic dashboard and trend visualization tools for a variety of customers.
For customers with more sophisticated needs or with various other types of data to be used in analysis, the camera metadata can be accessed and combined with other data and processed by other platforms for sophisticated visualization and data mining. This allows technology partners to access the data we aggregate into their own charts, graphs and exception reports powered by specialized software companies they may already be using.
There are familiar use cases spanning multiple industries that
WWW.SECURITYTODAY.COM 21