Let’s ‘AI’ the edge in video security!

AI on the edge is now increasingly employed in video security, where users can rely on AI video analytics in the camera to gain situational awareness and respond to events quickly. Yet, with increased computing power in the camera to support AI on the edge, memory and storage in the camera must also keep up with these evolving changes to ensure the camera runs at optimal stability and reliability. In this article, David Henderson, Director for Industrial/IMM for Micron Technology, speaks to asmag.com about key memory and storage elements AI camera manufacturers must consider, and what storage solutions Micron has to offer in this regard.

asmag.com: Why do we need AI at the edge in video security applications?

Henderson: AI implementation in video security applications should consider the factor of architecture from end-to-end, edge-to-cloud perspective. With the increasing processing power and algorithm development at edge devices in recent years, the AI-enabled camera can run advanced video analytics to extract valuable, actionable insights from captured data. By enabling AI functionality in the camera, the AI workloads can be split into camera devices and cloud data centers for operating efficiency, to provide a faster response (less latency) and to reduce bandwidth consumption. For example, in a situation where a security system needs instant response and action after analyzing a face or a number plate, sending data to the cloud and waiting for its answer is not feasible.

This is why edge analytics is becoming an area of significant investment for video security systems. It offers low transmission bandwidth consumption as only necessary data is sent to the cloud. Enabling AI at the camera ensures quicker alerts in case of threat detection, allowing faster data-driven analysis and decision-making. Edge-based analytics also come with lower hardware and deployment costs as less on-premises server resources are needed for the security solution.

asmag.com: Could you please share with us an example of edge AI usage? 

Henderson: These AI capabilities become even more essential as large-scale projects such as smart cities are growing and having more advanced functionality. Such trends lead to the development of new applications and AI-enabled camera solutions.

Multi-imager cameras
Multi-imager cameras, sometimes referred to as multi-sensor cameras, typically use multiple camera lenses to offer panoramic video coverage of 180 to 360-degree scenes. Multi-direction overviews combined with artificial intelligence make them ideal for wide-area coverage applications such as traffic intersections, retail and public spaces.
 
Capturing up to four independent video streams, each image sensor can be customized using unique video analytic parameters to detect only people and objects of interest per scene while automatically optimizing images. For retail, a loitering detection feature can provide alerts when people linger at a location for longer than usual. Facial recognition matched with watch lists can immediately detect people who may be known as threats.
 
Automatic number plate recognition (ANPR) cameras
The use of ANPR cameras has risen due to its wide range of applications in traffic management, smart parking, toll automation and intelligent transportation systems in smart cities. For business organizations, cameras with embedded AI-powered ANPR software can automatically recognize license plates and store the metadata in a database for future searches, while comparing them against watch lists to identify whitelisted, blacklisted or suspect vehicles to trigger actions such as opening a barrier gate for access control to restricted sites. 
 
Nowadays, smart cities can integrate ANPR cameras to city sensors for vehicle behavior analysis and traffic optimization, able to address free flow tolling and real-time road monitoring challenges. Automated parking systems and on-street and off-street ANPR parking solutions decrease the need for gates, tickets and operators and reduce traffic congestion.

asmag.com: From memory and storage point of view, what are the key elements of implementing AI at the camera that manufacturers need to consider?

Henderson: New video security cameras are now being designed to process multiple video streams and higher resolutions (4K and above) to provide AI algorithms with the large dataset of detailed images and videos required for them to analyze. In addition to this, increasing amounts of metadata are captured and stored on the device to enable operators to quickly search and find relevant video footage. Much of the processing now occurs on the device level. Implementing AI at the camera brings unique challenges as it may have power constraints, performance limits, durability issues and environmental impacts. Here are some factors that camera manufacturers should consider while designing their AI-enabled cameras.

• Applying low power, high performance memory solutions for data processing
• Utilizing high quality, high endurance, application-optimized storage solutions for code, application, and data storage
• Ensuring the memory and storage solutions support a wide temperature range for both indoor and outdoor use cases

asmag.com: What is Micron’s solution for AI-enabled camera designs?

Henderson: With more computing power from new chipsets enabling deep neural network processing on the camera itself for edge intelligence, memory and storage technologies need to keep up with these evolving changes in processing and workload requirements.
 
Micron’s memory and storage solutions for AI-enabled camera offer:

• LPDDR4 and LPDDR5 for edge computing

  • High-performance and low-power memory

  • Offer from 4Gb-128Gb densities

  • Support wide temperature for industrial use cases (-40°C to +95°C)

• e.MMC for code/application storage

  • ​High-capacity and high-performance storage

  • Offer from 32GB to 256GB densities

  • Support wide temperature for industrial use cases (-40°C to +85°C)

• Industrial microSD card for data storage

  • Storage solution optimized for AI workloads

  • Offer from 64GB to 1.5TB

  • Ability to concurrently handle 4K video recording and up to 8 AI event/second capturing

  • Provide a 2 million hour mean time to failure (MTTF) rating

  • Support wide temperature (-25°C to +85°C)