Scalable video for security camera platforms
What if cameras could intelligently feed AI and human pipelines only the pixels they actually need?
Security camera platforms need to support higher resolutions, more streams and growing AI workloads within limited compute, memory and power budgets.
MPEG-5 LCEVC combines lower-resolution base video with enhancement data for full-resolution reconstruction. Suitable analytics workloads can use the base video, while applications that need greater detail can use both layers.
The challenge
Different security camera applications need different levels of image detail. Some analytics can operate at lower resolution, while operators reviewing an incident may need the full-resolution picture.
In downstream pipelines this means decoding every frame at full resolution only to resize it before inference, adding to decoding work and costly memory requirements without necessarily improving the analytics result.
As stream counts and resolutions increase, this unnecessary processing leaves fewer resources available for other workloads.
The solution
MPEG-5 LCEVC works alongside existing codecs, including AVC, HEVC, AV1 and VVC.
The underlying codec encodes a lower-resolution version of the video, known as the base layer. A lightweight LCEVC enhancement layer carries the additional information needed to reconstruct the full-resolution image. This information can be combined with the base layer or carried separately.
This scalable approach means, for example, that analytics and AI inference pipelines can decode only the base layer, when its resolution meets the model’s input and accuracy requirements. Both layers can be used for full-resolution viewing, or retained for archival and future review.
Benefits
- Free resources for AI workloads: For analytics that do not require full-resolution input, the lower-resolution base representation can be processed directly. In a paper presented jointly with Intel, this delivered up to 3.4× higher throughput, approximately 70% lower power consumption and approximately 80% lower memory-bandwidth demand, while maintaining inference accuracy.
- Prioritize detail where it matters: The LCEVC enhancement layer can allocate additional detail selectively to regions of interest such as faces, licence plates or text, concentrating available bitrate where additional visual information has the greatest value.
- Reduce bandwidth and storage requirements: In MPEG verification tests, LCEVC delivered average bitrate savings of 28% for HD and 46% for UHD with AVC, and 24% for HD and 31% for UHD with HEVC, at comparable subjective visual quality. Across large camera deployments, lower bitrates can reduce both network traffic and storage requirements.
- Retain flexibility across codec generations: LCEVC can operate alongside AVC, HEVC, AV1 and other codecs, allowing manufacturers to apply the same enhancement approach as codec requirements evolve.
- Support more high-resolution streams: SOCs can design expensive codecs, such as AV1, to work at the lower resolution while applying LCEVC to get to the required high resolution. An illustrative camera-SoC model using AV1 indicates approximately 2.3× increase in 4Kp60 stream density.
Why it matters
Security platforms face resource pressure in two places: on the camera, where system-on-chip designs must encode more high-resolution streams within fixed compute and power budgets, and downstream, where analytics pipelines must decode and process a growing number of streams for AI inference.
LCEVC addresses both. Encoding the base layer at lower resolution reduces the load on the underlying codec, potentially increasing stream density per chip. A better compression performance also relieves bandwidth pressure on congested and costly networks. Downstream, suitable analytics can work directly from the base layer, avoiding unnecessary full-resolution decoding and freeing resources for more streams or additional workloads. Full-resolution video remains available for operators and investigations.
Explore
