Traffic Light Configuration Classification Under Suboptimal Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately determining the state of traffic lights in various environments, which affects their ability to control vehicle operations effectively, especially in situations where multiple configurations and suboptimal image conditions are encountered.
Innovation Solution
A method implemented by autonomous vehicle processors that captures images of traffic lights, processes them using machine learning classifiers based on pre-defined configurations, and selects the current state of the traffic lights to control the vehicle, involving image cropping, transformation, and the use of specific machine learning classifiers tailored to the traffic light configurations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single machine learning classifier is used to process all traffic light images, then the system complexity is low, but the accuracy of traffic light state prediction decreases in suboptimal image conditions and for diverse traffic light configurations
Solution Approach 1:
The patent segments the traffic light classification task by dividing it into multiple specialized machine learning classifiers, each trained on specific traffic light configurations (e.g., different orientations, bulb patterns, mounting styles). This segmentation allows each classifier to specialize in particular conditions, improving overall prediction accuracy across diverse and suboptimal scenarios while maintaining manageable system complexity through modular architecture.
Solution Approach 2:
The patent applies local quality by assigning different processing characteristics to different regions of the image based on detected traffic light features. The system identifies specific traffic light configurations in the image and routes them to appropriate specialized classifiers, allowing each region to be processed by the most suitable model for that particular traffic light type, thereby improving local and overall prediction accuracy.
2Measurement precision
If image processing is performed on the entire original image, then all traffic lights are captured, but the processing efficiency and accuracy for individual traffic lights decrease due to irrelevant background information
Solution Approach 1:
The patent extracts the relevant traffic light regions from the original image by detecting traffic light positions and configurations. It crops or segments the image to isolate individual traffic lights, removing irrelevant background information. This extraction process improves detection accuracy by focusing computational resources on the actual traffic light features while reducing processing time through smaller, targeted image regions.
3Adaptability or versatility
If the system uses multiple disparate traffic light configurations in a taxonomy, then the adaptability to different traffic light types improves, but the complexity of selecting and processing appropriate classifiers increases
Solution Approach 1:
The patent implements preliminary action by pre-organizing a taxonomy of traffic light configurations and training specialized classifiers for each configuration type before deployment. During operation, the system quickly matches detected traffic light features to the pre-established taxonomy and selects the corresponding pre-trained classifier, avoiding the need for complex real-time decisions and reducing processing complexity while maintaining high adaptability to diverse traffic light types.
Data Source
AI summary
Systems and methods for determining a current state, of at least one traffic light, for use in controlling an autonomous vehicle are described herein. Implementations determine, based on a pose instance of the autonomous vehicle and a stored mapping of an environment of the autonomous vehicle, a region of the environment includes the traffic light and a configuration that is assigned to the traffic light. Further, those implementations process an image capturing the region, using a machine learning classifier, to generate predicted output associated with multiple candidate states of the traffic light, and determine a current state of the traffic light based on the predicted output. Processing the image using the machine learning classifier can be based on the configuration of the traffic light.