PTAB

IPR2026-00108

Samsara Inc v. Motive Technologies Inc

Log In
Key Events
Petition
petition

1. Case Identification

2. Patent Overview

  • Title: Distracted Driver Detection
  • Brief Description: The ’243 patent describes a computer-implemented, multitask training pipeline for distracted-driver detection. The system uses a shared convolutional neural network (CNN) backbone to extract features from driver images, which are then fed into multiple task-specific prediction heads, including heads for distraction classification, object detection, and pose estimation.

3. Grounds for Unpatentability

Ground 1: Claims 1, 2, 6-14, 18-20, and 24-33 are obvious over Chun in view of Chao.

  • Prior Art Relied Upon: Chun ("NADS-Net: A nimble architecture for driver and seat belt detection via convolutional neural networks," 2019) and Chao (Application # 2020/0242381).
  • Core Argument for this Ground:
    • Prior Art Mapping: Petitioner argued that Chun taught the core architecture of the ’243 patent: a multi-task CNN (NADS-Net) with a shared feature pyramid network (FPN) backbone and multiple prediction heads for driver monitoring tasks, including pose estimation and seat-belt detection. However, Petitioner contended Chun did not explicitly name a separate distraction classification head. Chao was alleged to supply this missing element, as it disclosed a CNN system for monitoring vehicle occupants that included a distraction-classification head for predicting categorical states like "Drink," "Call," or "Text."
    • Motivation to Combine: A POSITA would combine Chun and Chao because both addressed the same problem of distracted driving detection using compatible, modular CNN architectures. Petitioner asserted a POSITA would be motivated to add Chao’s explicit distraction-classification head to Chun’s multi-head framework to convert the features Chun already extracted (e.g., for drinking or phone use) into a discrete classification output, which was described as the natural next step in solving the problem Chun identified.
    • Expectation of Success: Success was expected because adding a classification head to a shared backbone was a routine multi-task extension. Chun’s backbone was already proven to drive multiple heads, and Chao's head was designed to consume the same type of features (related to face, hand, and body position) that Chun’s backbone already extracted.

Ground 2: Claims 3-5, 10, 15-17, and 21-23 are obvious over Chun in view of Chao and further in view of Shanmugamani.

  • Prior Art Relied Upon: Chun, Chao, and Shanmugamani ("Deep Learning for Computer Vision," 2018).
  • Core Argument for this Ground:
    • Prior Art Mapping: This ground built upon the Chun and Chao combination to address dependent claims reciting specific internal structures of the prediction heads. Petitioner argued that while Chun and Chao taught the high-level components, Shanmugamani provided the explicit, conventional implementation details. Shanmugamani was presented as a deep-learning textbook that taught standard object detection patterns, such as using a bounding-box regression network alongside a class-prediction network (claim 3), and implementing these networks with hidden layers containing convolutional, batch normalization, and activation layers (claims 4 and 5).
    • Motivation to Combine: A POSITA implementing the system of Chun (which used the TensorFlow framework) would naturally consult a contemporaneous reference like Shanmugamani for standard TensorFlow layer patterns and implementation recipes. Shanmugamani provided the "off-the-shelf" components for building the object detection and pose estimation heads that Chun and Chao required.
    • Expectation of Success: Success was expected because Shanmugamani was an implementation guide for the exact type of automotive computer vision tasks at issue. Applying its standard TensorFlow blocks to Chun’s TensorFlow-based model was described as a routine and predictable substitution.

Ground 3: Claims 1-6, 9-18, and 20-24 are obvious over Zhengyang in view of He.

  • Prior Art Relied Upon: Zhengyang (Chinese Publication # CN111860253) and He ("Mask R-CNN," 2017).

  • Core Argument for this Ground:

    • Prior Art Mapping: Petitioner argued that Zhengyang disclosed a multi-task CNN for driver-attribute recognition with a shared backbone (SE-ResNet18) and multiple prediction heads that performed distraction classification (e.g., "smoking recognition," "phone call recognition") and object detection. However, Zhengyang did not explicitly teach a pose-estimation head. He was alleged to provide this element, teaching a modular keypoint head for human pose estimation that could be readily added to a shared backbone framework like Mask R-CNN, operating in parallel with other heads.
    • Motivation to Combine: A POSITA would combine Zhengyang and He to improve the accuracy and comprehensiveness of Zhengyang’s driver-state assessment. Zhengyang’s goal was to "comprehensively reflect the driver's driving state," and adding He's pose-estimation head would provide crucial data on head, arm, and body orientation—highly discriminative features for distraction detection. The modular nature of He's keypoint head made it suitable for integration into Zhengyang's multi-head architecture.
    • Expectation of Success: A POSITA would expect success because both references used standard deep learning building blocks (e.g., ResNet-style backbones, convolutional heads) and shared-feature architectures. Adding a pose estimation head was a well-understood extension for multi-task models by the patent’s priority date.

  • Additional Grounds: Petitioner asserted an additional obviousness challenge (Ground 4) based on the combination of Zhengyang, He, and Chao for claims 7-8, 19, 25-33. This ground primarily argued that Chao supplied explicit teachings for taking action (e.g., alerts) based on the distraction tags generated by the Zhengyang and He system.

4. Relief Requested

  • Petitioner requests institution of an inter partes review and cancellation of claims 1-33 of the ’243 patent as unpatentable.