Machine-Driven Video Prompts Industry to Rethink Compression, Computing and Network Architecture

WASHINGTON, DC, UNITED STATES, October 29, 2026 /EINPresswire.com/ — Video infrastructure is approaching an architectural turning point as artificial intelligence applications increasingly rely on cameras and visual sensors to interpret physical environments, coordinate activities and make decisions without human intervention.

The shift is exposing limitations in the scalability of physical AI due to currently-available video compression technologies, which were developed to deliver high-quality images to human viewers. Applications ranging from autonomous vehicles and industrial robotics to traffic management and infrastructure inspection require visual information to be processed continuously, often under strict latency, bandwidth and power constraints. Whereas visual fidelity is critical for humans viewing video, one of these machine use cases processing video needs only visual information sufficient for it to perform its function, with any additional visual information potentially wasting bandwidth and power and even impairing the machines’ function.

These requirements are prompting the video technology industry to reconsider how visual information is captured, compressed, transmitted and processed. Emerging approaches, including Feature Coding for Machines (FCM) and Versatile Video Coding for Machines (VVC for Machines), could substantially change the economics and architecture of machine vision systems.

During a recent executive vidcast for journalists featuring members of the Media Coding Industry Forum (MC-IF), Aytac Biber of Qualcomm, Honglei Zhang of Nokia and Rob Moore of OP Solutions examined how machine-oriented video compression could address the operational demands of physical AI.

The discussion highlighted a fundamental change in the industry’s priorities. Compression efficiency increasingly must be evaluated against the accuracy of machine decisions, the distribution of computational workloads and the cost of moving visual information across networks.

Machine Vision Changes the Economics of Video Compression

Conventional video codecs have evolved around the requirements of human consumption. Standards such as MPEG-2, Advanced Video Coding (AVC) and High Efficiency Video Coding (HEVC) improved the efficiency of storing and distributing video while maintaining image quality suitable for viewers. Machine vision applications introduce different requirements.

Biber explained that industrial cameras, autonomous vehicles, drones and robotic systems increasingly capture visual information to identify objects, track movement, recognize patterns and support automated decisions. The visual characteristics that matter to human viewers may have little relevance to these applications.

An autonomous vehicle, for example, must identify pedestrians, road boundaries, obstacles and other vehicles. Preserving the visual richness of surrounding scenery may contribute little to the vehicle’s immediate decision-making requirements.

Moore emphasized that conventional codecs can therefore preserve substantial amounts of information that machine learning systems subsequently discard during analysis. The consequences extend beyond bandwidth consumption. Conventional architectures typically require cameras to encode visual information, transmit compressed video to a processor, decode the images and then apply neural networks to extract the features necessary for inference. This sequence can create processing bottlenecks, such as when multiple cameras continuously feed a centralized computing system.

FCM approaches the compression differently. Physical use cases rely on features, machine-readable representations of information extracted from the video, rather than pixel information. Whereas most physical AI systems implemented using traditional video codecs extract features from the pixel information of a decoded video bitstream, FCM shifts the extraction process to the edge and produces a bitstream consisting of features alone. FCM bitstreams use significantly less bandwidth and otherwise perform better than traditional codecs for many physical AI implementations. Moore cited performance comparisons in which FCM achieved equivalent object detection and tracking accuracy compared to traditional codecs while using 18 to 25 times less bandwidth. These figures illustrate the potential efficiency gains, although actual performance depends on the application, configuration and evaluation conditions.

Zhang noted that machine vision also requires a broader approach to measuring compression performance. Different applications have different accuracy requirements, and seemingly minor errors in visual interpretation can have serious operational consequences. The relevant performance measures consequently extend beyond image quality to include detection accuracy, processing efficiency and the reliability of downstream machine decisions.

Physical AI Pushes Processing Toward the Edge

The expansion of physical AI could place substantial additional demands on enterprise networks, local computing infrastructure and semiconductor platforms.

Continuous video capture creates particularly difficult scaling challenges. Unlike conventional entertainment video, machine-generated visual information may be produced around the clock by thousands of cameras, sensors and autonomous devices. Moore illustrated the potential scale using autonomous vehicles equipped with approximately 17 cameras. He estimated that such a vehicle could generate roughly 5.5 terabytes of raw video information per hour, creating enormous aggregate data volumes as autonomous transportation expands.

These estimates represent potential data generation rather than traffic necessarily transmitted across external networks. Nevertheless, they underscore the importance of reducing unnecessary data movement.

Biber observed that the infrastructure implications extend well beyond wide-area networks. Local networks, edge servers, memory systems, storage resources and power supplies must also accommodate increasing volumes of machine-generated information. Machine-oriented compression could help address these pressures by enabling more processing to occur near the cameras and sensors generating the data. The implications are particularly significant for semiconductor design.

According to Biber, machine vision is becoming an increasingly important consideration in how chipmakers integrate camera processing, AI accelerators, memory, video engines, connectivity and security capabilities.
Efficient feature representation could reduce memory transfers and network traffic while allowing portions of AI inference workloads to be distributed among devices, edge processors and centralized computing systems.

Moore also described a potential evolution beyond conventional device-to-server architectures toward direct communication among intelligent edge devices. Autonomous vehicles, for example, could exchange information about road hazards without transmitting complete video streams. Similarly, coordinated industrial robots could share relevant observations to improve their collective understanding of an operating environment. Such architectures could reduce dependence on centralized processing while introducing new requirements for interoperability, synchronization and reliable communication.

Privacy represents another important consideration. Biber explained that transmitting extracted features rather than complete images could reduce the amount of sensitive visual information leaving a device. However, he cautioned that feature-based compression does not independently guarantee privacy. Security controls, system architecture and the possibility of reconstructing sensitive information from transmitted features remain essential considerations.

Competing Compression Approaches Create New Implementation Choices

As the machine vision market emerges there is also the potential to support multiple compression architectures.

Zhang distinguished between VVC for Machines and FCM, explaining that the approaches offer different implementation advantages. VVC-based techniques retain a conventional video representation while incorporating methods that improve performance for machine analysis. Neural networks can support preprocessing and post-processing to optimize the resulting information for specific tasks. This approach can provide a more familiar migration path for organizations with existing video infrastructure, although reconstructing image data can preserve some of the processing overhead associated with conventional codecs.

FCM takes a more specialized approach by compressing intermediate neural network features for consumption by downstream machine learning systems. That architecture can provide greater compression efficiency and flexibility in distributing computational workloads, but implementation depends more directly on the requirements of the underlying AI models and applications.

For enterprise technology leaders, the choice involves trade-offs among infrastructure compatibility, bandwidth efficiency, computational requirements, accuracy and deployment complexity. Zhang said the industry has already made meaningful progress toward formalizing these approaches. Technical work supporting VVC for Machines is available, while FCM standardization has reached an advanced stage.

Moore nevertheless emphasized that further collaboration and implementation feedback will be necessary as machine vision requirements evolve.

The broader significance extends beyond improvements in compression ratios. As visual sensors become integral to industrial automation and autonomous systems, the representation of visual information will increasingly influence network design, semiconductor architecture and the distribution of AI workloads. Organizations evaluating machine vision infrastructure will consequently need to consider how compression decisions affect the performance and economics of the entire system.

The industry’s emerging challenge is to develop interoperable approaches that preserve the information machines require while minimizing the computational and network resources consumed in delivering it. For physical AI, that capability could become an important determinant of how efficiently intelligent systems scale from individual devices to coordinated networks of autonomous machines.

Click here to read the Q&A based on this interview.

Airrion Andrews
Mindshare Capture

Media gallery