HD Mobile Video Surveillance: Surpassing Network Limitations
Intel-based network appliances offer power densities that can help unburden your core networking infrastructure.
Many companies are just now starting to upgrade their video surveillance infrastructure to high definition (1 Mega Pixel +), and with the technology growing cheaper by the minute, many more will soon be upgrading to the latest and greatest technology. This jump in quality doesn’t just require upgraded cameras, it also needs significant supporting network/storage infrastructure. The amount of IoT devices is set to grow exponentially over the following years, and video surveillance is by far one of the most bandwidth-hungry applications, with equally large storage requirements.
Better Resolution = Better Surveillance
Figure 1, showing an HD (720p) video surveillance screenshot and an SD (480P) still, illustrates how a 4x increase in resolution brings a near-equivalent increase in detail. Given this, it’s easy to understand why a high-resolution camera (2MP or 1080p) is ideal for implementing useful analytics like facial recognition. Once we reach 4K resolution the amount of detail will open up even more applications. Small text on nametags, hand gestures, and even lip-reading can become viable.
The increase in overall detail can help improve context, make it possible to discern smaller objects, and provide a solid evidence-admissible recording for liability reasons.
Struggling to Scale for HD Video
Just a few 4k video cameras are enough to saturate the bandwidth of modern Internet connections, and even enterprise cloud service providers are hard-pressed to make cloud-based video analytics available on streams greater than 480p(SD). Even with the impending upgrades to telecom infrastructure, centralized systems simply cannot be cost-effective in the face of hundreds/thousands of bandwidth-saturating devices.
Table 1 shows costs for a single day of storage, but keep in mind that recordings are typically stored for much longer than that. What’s more, network and supporting infrastructure costs aren’t even factored into this equation. These are the main reasons video-streaming giants like Netflix rely heavily on a more distributed model by using content delivery networks (CDNs). Add in the fact that in-vehicle and rolling stock surveillance is moving to Solid State Drives (for reliability reasons) while running on less capable wireless technologies, and things start getting expensive fast.
Network Video Recorders (NVRs) began with one simple function, essentially a next-gen DVR for IP-based video cameras. But its optimal position close to the video source has been increasingly acknowledged and exploited by industry players. Today’s most sophisticated systems are essentially purpose-built network appliances. One example of such a system, running specialized Intel® x86 video surveillance software, is shown in Figure 2.
X86-based appliances bring with them the major advantage of an extensive ecosystem of proprietary and open source software/libraries that make it the platform of choice in most cases.
The Secret Sauce: Powerful Analytics at the Edge
To avoid costly infrastructure and the embarrassment of DDoS’ing your own video surveillance infrastructure, you need to circumvent the bottleneck that is the round trip to the data center. To accomplish this, a growing trend is spanning across all industries which are tightly ingrained with information technology: Edge Computing.
By harnessing the great power-densities of Intel®-based network appliances and implementing efficient edge computing techniques combined with analytics, solution providers and even companies rolling their own systems have been creating highly scalable, cost-effective solutions. At the same time, powerful machine-learning algorithms at the edge are giving viability to many lucrative and beneficial applications.
Achieving Decentralized Analytics
Video-streams, even highly compressed ones, are by far the largest consumers of Internet bandwidth. There is no easy way around the heavy requirements of video-streaming aside from the incremental improvements encoders/decoders and specialized codecs like H.264/H.265 provide. But there is a way to strip useful information from all the noise in HD video streams and minimize bandwidth usage: Pre-process video streams as close to the source as possible.
In a world where Big Data continues to steamroll the opposition, metadata is king. Metadata is used by Google, for example, and is more or less the concentrated secret sauce of the data-driven businesses of today. Born from the need to efficiently classify, store, and transmit large amounts of information, metadata is data which describes another dataset.
Converting to Text-Based Metadata
Imagine you’re law-enforcement and searching the feed for a suspect based on appearance. How would you go about it?
You could narrow down the search criteria by area (i.e., camera feed), but what about searching for people with a red hoodie? Or a license plate number? Certain hair color or facial features? An object travelling at xx speed? Such information is what an edge analytics-driven system would continuously transmit to the data center, instead of entire video feeds. Just as important, it would transmit unique and distinguishing data that is immediately useful for analytics. The techniques and algorithms involved in this process are highly sophisticated and still reaching maturity, as video feed information presents extremely noisy and unstructured data.
Modern Network video recorders house enough storage for several days, even weeks of HD video surveillance. This is enough to comply with law enforcement and more than enough for 90 percent of use cases. Video Management Systems (VMS) can intelligently manage recordings of importance in permanent storage based on analyzed metadata.
If cameras stream live feed to the operator room anyway, why not just centralize storage in that case? As seen in Table 1, there is a large difference in streaming SD, HD and 4k media. With the amount of bandwidth that’s needed to transmit the feeds from a couple of HD cameras, you could easily stream many times that amount of low-quality feeds. While much lower in detail and definition, such feeds convey the benefit of allowing all streams to be viewed simultaneously. It’s possible to look for the target and pinpoint the feed that needs to be accessed in full definition for further inspection. Now instead of needing core networking infrastructure capable of handling the full load of all the HD Cameras, you can effectively get away with 1/10th of the bandwidth. This technique is exploited in most modern NVRs, which all incorporate the capability to stream low-quality feeds and store high-quality versions locally.
More than Just Video Surveillance Moves Computing Closer to the Edge
IoT and mobile phones have ushered in a staggering amount of bandwidth-consuming devices, and exponentially more are on the way. Decentralized computing models are the only ones capable of keeping up with the ever-growing demands, and major players (Verizon, AT&T, the largest proponents of 5G and MEC) have already began morphing their networks in preparation for the upcoming challenges. We will be seeing less fallback/reliance on the cloud as businesses move to more scalable solutions.
James Piedra is a Network Platform Analyst at Lanner Electronics (lanner-america.com), a leading designer and manufacturer of Embedded Computers and Network Appliances. He researches and writes about the latest advances in Information technology, mainly focusing on Software-defined Networking, IoT, Cyber Security and mobile. On his free time James likes to tinker with consumer/embedded electronics and open source software (most time spent fixing something he broke in the process).