Object Detection Signatures With Error-Resolving AI Routing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing perception systems in ADAS and AVs face challenges in achieving the required accuracy of 0.999 due to the limitations of gradient descent-based deep learning algorithms, which fail to address rare edge cases and require excessive computational resources.
Innovation Solution
An adaptable AI system that includes an error resolving part (ERP) to dynamically adapt to errors by generating accurate signatures, utilizing cortical units and error resolving units to correct erroneous signatures, thereby enhancing object detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deeper and heavier neural networks are used to improve detection accuracy, then accuracy improves, but computational power requirements increase by over x10
Solution Approach 1:
The patent segments the detection task by creating a hierarchical network structure with a first neural network for initial detection and a second neural network for verifying specific object classes. This segmentation allows the system to achieve high accuracy for critical objects without requiring all detectors to operate at maximum depth simultaneously, thus reducing overall computational power requirements while maintaining detection accuracy above 0.999 for specified objects
Solution Approach 2:
The patent applies local quality by configuring different neural networks with different depths and complexities based on their specific detection tasks. The second neural network is configured with greater depth specifically for detecting specified objects where high accuracy is critical, while other detection tasks use shallower networks. This localized optimization achieves high accuracy where needed without proportionally increasing power consumption across the entire system
2Measurement precision
If retraining with more labeled data is performed to improve accuracy, then some false-positives/false-negatives are solved, but new false-positives/false negatives are exposed and accuracy saturation is reached
Solution Approach 1:
The patent implements preliminary action by pre-training the second neural network offline with labeled data for specified object classes before deployment. This preliminary training allows the system to achieve high accuracy for critical objects without requiring continuous online retraining. The pre-trained second network consistently verifies detections with high reliability, avoiding the accuracy saturation and inconsistency problems that occur with iterative retraining approaches
Data Source
AI summary
A method that is computer implemented and is for solving inaccuracies associated with object detection, the method includes automatically evaluating, by a controller, an accuracy of signatures for use in the object detection, the signatures were generated by an adaptable artificial intelligence (AI) system. When finding an erroneous signature of the signatures, by the controller, triggering a generation of (a) a narrow AI agent configured to solve an error associated with the erroneous signature and (b) a router that routes a sensed information unit (SIU) associated with the erroneous signature to the narrow AI agent, wherein the erroneous signature, when used for the object detection, results in an object detection error.