A Computer-Implemented Method for Segmenting Patterns in a Sequence of Frames

Simple SummaryContent extracted from patent full text and abstract with AI.

This patent discloses a computer-implemented method for segmenting patterns, such as actions or events, in a sequence of frames (like video frames, audio samples, images, or sensor data). The core innovation is an efficient algorithm that computes, for each frame, the probability of different patterns, and then iteratively refines the labeling of these frames using a differentiable approximation of a complex, originally non-differentiable energy function. By making the process differentiable and amenable to modern optimization techniques (like gradient descent), the method allows faster, more robust, and more parallelizable pattern segmentation, making it especially suitable for real-time applications.

Use CasesContent extracted from patent full text and abstract with AI.

• Action recognition and segmentation in video surveillance systems (e.g., detecting specific activities or events in video footage).
• Real-time object or scene segmentation for autonomous vehicles (e.g., interpreting sensor or camera data to detect road events).
• Quality control and activity monitoring on manufacturing lines using visual sensors (e.g., detecting production steps or faults).
• Audio event detection and segmentation in security or smart assistant systems (e.g., detecting and classifying sounds or speakers over time).
• Medical signal analysis, such as segmenting patterns in EEG or ECG data for disease diagnosis or monitoring.
• Satellite or aerial image strip analysis for land use or anomaly detection.

BenefitsContent extracted from patent full text and abstract with AI.

• Enables fast and efficient segmentation of sequential data, suitable for real-time applications.
• Greater parallelizability, allowing for deployment on modern multi-core or GPU systems.
• Reduces computational cost compared to traditional dynamic programming approaches, with linear rather than quadratic time complexity with respect to sequence length.
• More robust results even with noisy initializations, as the approach improves reliability and segmentation quality.
• Highly flexible, able to work with different modalities (images, audio, sensor data) and various types of segmentation tasks (action, object, audio event, etc.).
• Allows a trade-off between speed and accuracy by adjusting optimization parameters.
• Can be integrated into existing machine learning frameworks with ease due to differentiable formulation.

Technical Classifications (CPCs)

Inventors & Applicants

Patent Abstract

A computer-implemented method for segmenting patterns in a sequence of frames over a succession dimension, the method comprising:- computing frame-wise pattern probability estimates (26) for the sequence;- inferring, from the frame-wise pattern probability estimates (26) and a set of possible patterns, predicted frame-wise pattern labels (32) for the sequence, wherein the inferring comprises iteratively optimizing a differentiable approximation (E∗) of an energy function, the energy function being a non-differentiable function that measures a compatibility between the frame-wise pattern probability estimates (26), the set of possible patterns and the predicted frame-wise pattern labels in the sequence.