Hayden AI Technologies v. Safe Fleet — Court Defines Mobile Edge and Deep-Learning Terms in Traffic-Enforcement Patents

Background

Hayden AI Technologies accused Safe Fleet and related companies of infringing patents covering systems that use mobile detection devices to identify traffic violations. The technology processes video from vehicles, identifies vehicles and restricted road areas, and produces evidence packages. One patent describes lane-violation detection using convolutional neural networks, placing the dispute at the intersection of patent law, computer vision, and edge artificial intelligence.

After a Markman hearing, Magistrate Judge James Cho construed eleven disputed claim terms from U.S. Patent Nos. 11,003,919 and 11,164,014. Claim construction determines the legal meaning of patent language and will shape the later infringement and validity analysis; the order did not decide who ultimately wins the case.

The Court’s Holding

The court construed “edge device” as a mobile edge device. The claims and specification repeatedly tied the device to cameras and processing systems carried by vehicles, so the broader meaning Hayden proposed was not appropriate. It defined a “bounding box” as a rectangular or quadrilateral shape enclosing a detected object, rejecting a definition confined to rectangles.

Several software terms kept their ordinary meaning. “Vehicle attributes,” “computer vision library,” “plurality of functions,” and “docker container” did not require the narrower technical limitations defendants proposed. By contrast, the court gave “docker container image” a detailed meaning: a lightweight, standalone, executable package containing everything needed to run software or read or manipulate data, including code, runtime instructions, tools, libraries, and settings.

The most consequential construction concerned “bounding” vehicles and restricted road areas. Although the claims’ grammar did not expressly say the deep-learning model performed every bounding step, the specification consistently described the model applying bounding boxes to detected objects. Reading the claims in the context of the patent as a whole, the court therefore required the bounding boxes to be output by the deep-learning model.

Key Takeaways

• Patent terms used in edge-AI systems are interpreted through the entire patent, not modern industry usage viewed in isolation.
• Repeated descriptions of a mobile implementation can narrow a seemingly broad term such as “edge device.”
• Courts resist importing detailed functional limits into ordinary software terms unless the intrinsic record justifies them.
• A specification that consistently assigns a task to a deep-learning model can limit claims even when the claim language does not expressly repeat that assignment.

Why It Matters

The order illustrates how conventional claim-construction rules apply to AI-enabled inventions. Labels such as “deep learning,” “computer vision library,” and “edge device” do not decide scope by themselves. Drafting choices about architecture, mobility, and which component performs a function can become decisive years later in litigation.

For companies building automated enforcement, robotics, or vision systems, the ruling underscores the value of describing multiple implementations when seeking patents. For the parties, requiring deep-learning-generated bounding boxes may become central when Hayden compares the asserted claims with Safe Fleet’s accused system.