Sign in Subscribe
Machine learning

Adaptive Corner Detection Method for Machine Vision

Strategic Insights

— 2 min read

A recent patent application by Arindam Basu and Pao-Sheng Vincent Sun proposes a novel adaptive corner detection method for machine vision, addressing the challenges of efficient corner detection on resource-constrained edge devices. The method, named LuvHarris, leverages the Threshold-Ordinance Surface (TOS) data structure and a multi-threaded processing pipeline to achieve high accuracy and increased throughput. However, the inventors acknowledge the limitations of the current implementation, which struggles with energy efficiency on extreme-edge Application-Specific Integrated Circuits (ASICs). The proposed method aims to overcome these challenges by providing an adaptive corner detection technique that can be deployed on edge devices, utilizing data from dynamic vision sensors.

Key Takeaways:

  • The LuvHarris method is designed to address the limitations of current corner detection techniques, which struggle with throughput and accuracy on edge devices.
  • The proposed method uses the Threshold-Ordinance Surface (TOS) data structure and a multi-threaded processing pipeline to enhance processing speed while maintaining accuracy.
  • The LuvHarris method includes several key components, such as capturing and organizing event data into a 2D array, transforming the array into Ordered Surface matrices, and applying a corner detector to the OS matrices.
  • The method also includes embodiments for arranging recorded events in global temporal ordering, applying spatial-temporal correlation filters, and performing sort normalization to improve corner detection accuracy.
  • The proposed system includes a dynamic vision sensor and a machine vision processor, which are electrically connected to capture and process event data for corner detection.
  • The inventors claim the advantages of the method and system include efficient corner detection, high-speed processing, and low energy consumption.

Statistics:

  • The LuvHarris method is designed to process event data at high speeds, with the maximum number of recorded events in every row ranging from 25 to 100.
  • The method includes several embodiments, each with its own set of features and advantages:

+ Embodiment 1: Obtaining event data from a dynamic vision sensor and arranging recorded events in global temporal ordering.

+ Embodiment 2: Arranging recorded events in a plurality of rows in the 2D array.

+ Embodiment 3: Applying a spatial-temporal correlation filter to the event data before organizing the recorded events into the 2D array.

+ Embodiment 4: Performing sort normalization to the patch of elements before employing the Harris detector.

  • The method also includes several claims, which detail the specific features and advantages of the proposed invention:

+ Claim 1: A method for adaptive corner detection in machine vision.

+ Claim 2: Arranging recorded events in global temporal ordering in the 2D array.

+ Claim 3: Arranging recorded events in a plurality of rows in the 2D array.

+ Claim 4: Applying a spatial-temporal correlation filter to the event data before organizing the recorded events into the 2D array.

Sources:

  • BASU, Arindam; SUN, Pao-Sheng Vincent. Method And System For Adaptive Corner Detection Using Dynamic Vision Sensors. U.S. Patent Application Number 20250245958, filed January 31, 2024, and posted July 31, 2025.
  • Patent URL (for desktop use only): https://ppubs.uspto.gov/pubwebapp/external.html?q=(20250245958)&db=US-PGPUB&type=ids

Read more