Autonomous Driving Perception Using Predictive Narrow AI Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous vehicle perception systems are computationally intensive and power-consuming due to their need to exhaustively analyze and label raw sensor inputs to identify various objects and entities in the scene, which is inefficient and resource-heavy.
Innovation Solution
Implementing a perception unit that uses an ensemble of narrow AI agents, where only relevant agents are activated based on high-level contextual cues, reducing the need for detailed object detection and localization, thereby lowering power consumption and computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If exhaustive object detection and localization is performed to identify all objects and entities in the scene, then measurement precision and reliability are improved, but use of energy and computational resources worsen
Solution Approach 1:
The patent segments the scene into multiple regions of interest (ROIs) based on semantic information from a lightweight segmentation model. Instead of applying heavy detection models to the entire image, only selected ROIs are processed by object detection models. This segmentation approach maintains detection precision for relevant objects while significantly reducing computational load and power consumption by limiting processing to necessary areas.
Solution Approach 2:
The system performs partial object detection by selectively applying detection models only to regions where objects are likely to be present, rather than exhaustively scanning the entire scene. The selective ROI processing approach implements partial action - detecting only the subset of objects that are relevant to driving decisions, thereby reducing energy consumption while maintaining sufficient measurement precision for safety-critical objects.
2Measurement precision
If exhaustive object detection and localization is performed to identify all objects and entities in the scene, then measurement precision and reliability are improved, but computational intensity worsens
Solution Approach 1:
The patent divides the computational task into two stages: first, a lightweight segmentation model identifies regions of interest; second, heavy object detection models are applied only to these segmented ROIs. This segmentation strategy reduces computational intensity by avoiding redundant processing in empty or irrelevant regions, while maintaining detection precision through focused analysis of critical areas.
Solution Approach 2:
The system implements partial object detection by processing only a subset of the image (the ROIs) with computationally intensive detection models. This partial action approach reduces overall computational intensity while maintaining sufficient detection precision for safety-critical objects, avoiding the excessive computation required for exhaustive full-scene analysis.
3Reliability
If detailed object detection and localization is performed for all objects, then reliability of driving decisions is improved, but use of energy worsens
Solution Approach 1:
The patent segments the visual scene into regions of interest based on semantic segmentation, then applies object detection only to these segments. This ensures reliable detection of relevant objects (pedestrians, vehicles, obstacles) while avoiding energy-wasting processing of irrelevant areas, thus maintaining driving decision reliability with reduced power consumption.
Solution Approach 2:
The system performs partial detection by focusing computational resources only on regions where detection is necessary for safe driving. This partial action maintains reliability for critical driving decisions by thoroughly detecting relevant objects, while reducing energy consumption by skipping unnecessary detection in irrelevant regions.
Data Source
AI summary
A method for managing a group of narrow artificial intelligence (AI) agents for at least partially autonomous driving, the method includes (i) obtaining, by a prediction circuit, a stream of metadata segments generated at multiple points in time and associated with a selection of one or more sub-groups of the group of narrow AI agents; wherein the metadata segments are selected out of (a) selected narrow AI agent identifiers, and (b) multiple multi-domain identifier, the multiple multi-domain identifier are indicative of multiple instances of multi-domain information about elements affecting a vehicle in relation to the multiple points in time; wherein a multi-domain identifier generated at a given point of time of the multiple points in time is a combination of class signatures that are indicative of classes of elements of a multi-domain information associated with the given point in time; (ii) finding, by the prediction circuit, a segment of the stream that is a predictor to a receiving of a next cluster identifier at a future point in time; and (iii) automatically predicting, when finding the predictor, at least one of: (c) future metadata segments to be received during the future point of time, or (d) a future sub-group of narrow AI agents to be selected at the future point of time.