Machine Learning Models Operating At Different Frequencies For Autonomous Vehicles
Patent No. US11816585 (titled "Machine Learning Models Operating At Different Frequencies For Autonomous Vehicles") was filed by Tesla Inc on Dec 3, 2019.
What is this patent about?
’585 is related to the field of machine vision, specifically object detection for autonomous vehicles. Modern vehicles use cameras to perceive their surroundings, but the high frame rates of these cameras often outpace the processing capabilities of accurate deep learning models. This discrepancy leads to either underutilization of available visual data or a compromise in the accuracy of object detection, which is crucial for safe autonomous navigation.
The underlying idea behind ’585 is to leverage two machine learning models operating at different frequencies to achieve both speed and accuracy in object detection. A faster, but potentially less accurate, model processes images at the camera's full frame rate. A slower, more accurate model analyzes a subset of these images and periodically provides its output to the faster model, effectively recalibrating the faster model's performance.
The claims of ’585 focus on a method, system, and storage medium for processing images from vehicle-mounted cameras. The core of the invention involves analyzing images using a first machine learning model at the threshold frequency of the camera, while a second machine learning model analyzes a subset of the images at a lower frequency . The first model periodically receives output from the second, improving its accuracy in determining the location of classified objects.
In practice, this system allows for real-time object detection at the camera's frame rate. The faster model provides immediate, though potentially less precise, location information. The slower model, running in parallel, refines this information periodically. For example, the faster model might initially identify a pedestrian with a bounding box that is slightly off. When the slower model processes the same scene, it provides a more accurate bounding box, which is then used to correct the faster model's subsequent estimations.
This approach differentiates itself from prior solutions by intelligently combining the strengths of two models with different performance characteristics. Instead of relying solely on a fast, but inaccurate, model or being limited by the processing speed of a highly accurate model, ’585 achieves a balance. The periodic recalibration mechanism ensures that the faster model's accuracy is continuously improved, leading to more reliable object detection for autonomous driving.
How does this patent fit in bigger picture?
Technical landscape at the time
In the late 2010s when ’585 was filed, autonomous vehicle systems commonly relied on camera-based object detection, at a time when deep learning models were increasingly used but hardware or software constraints made real-time processing of high-resolution video feeds non-trivial. At that time, it was typical to balance model accuracy with processing speed, often requiring trade-offs between frame rate and the complexity of the machine learning algorithms employed.
Novelty and Inventive Step
The examiner approved the patent because the prior art, when considered as a whole, does not teach or suggest a method for operating an autonomous vehicle that involves obtaining images at a threshold frequency, analyzing a subset of those images with a second machine learning model at a lower frequency, and using the output of the second model to improve the accuracy of the first model in determining the location of objects. The examiner considered this combination of steps, particularly the underlined portions, to be non-obvious.
Claims
This patent contains 18 claims, of which claims 1, 12, and 16 are independent. The independent claims are directed to a method, a system, and computer storage media, respectively, all generally focused on determining location information of objects using a combination of machine learning models for autonomous driving. The dependent claims generally elaborate on the specific configurations and functionalities of the machine learning models and their interactions.
Key Claim Terms New
Definitions of key terms used in the patent claims.
Litigation Cases New
US Latest litigation cases involving this patent.
Patent Family
File Wrapper
The dossier documents provide a comprehensive record of the patent's prosecution history - including filings, correspondence, and decisions made by patent offices - and are crucial for understanding the patent's legal journey and any challenges it may have faced during examination.
Date
Description
Get instant alerts for new documents
US11816585
- Application Number
- US16701669
- Filing Date
- Dec 3, 2019
- Status
- Granted
- Expiry Date
- Dec 11, 2041
- External Links
- Slate, USPTO, Google Patents