Traffic Light Obstruction Learning With Augmented Occluder Crops
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately classifying traffic lights due to temporary obstructions, leading to incorrect classifications or no classification when the crop of captured images does not include the traffic light, as the classifier is confused by untrained image data.
Innovation Solution
A data augmentation process is employed to generate training data by acquiring and augmenting images of obstructing objects, such as large vehicles, by randomly rescaling and cropping them at random locations, which are then used to train neural networks to predict the state of traffic lights, whether obstructed or not, and handle 'unknown' states when obstructions are present.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the classifier is trained only on clear traffic light images, then classification accuracy for unobstructed traffic lights is improved, but classification reliability deteriorates when traffic lights are temporarily obstructed
Solution Approach 1:
The system performs preliminary actions by capturing multiple images of the same scene at different time points before making a classification decision. When a traffic light is obstructed in one image, the system has already captured or will capture additional images that may show the unobstructed traffic light, allowing the classifier to reliably determine the correct state despite temporary obstructions.
Solution Approach 2:
The system implements feedback by using classification results and obstruction detection outcomes to guide subsequent image capture decisions. When an obstruction is detected, the system triggers additional image captures and uses the results to refine the final classification, creating a closed-loop system that improves reliability through iterative verification.
2Reliability
If multiple images are captured and processed to handle obstructions, then classification reliability is improved, but processing time and computational complexity increase
Solution Approach 1:
The system applies partial action by processing only the necessary subset of captured images rather than all images uniformly. When a clear traffic light is detected in any image, the system can stop processing additional images for that particular classification task, reducing unnecessary processing time while maintaining reliability through selective image evaluation.
Solution Approach 2:
The system performs preliminary obstruction detection on captured images before full classification processing. By quickly identifying whether an image contains an obstruction, the system can prioritize processing of clear images and avoid wasting computational resources on obviously obstructed images, thereby reducing overall processing time.
Data Source
AI summary
Disclosed is data augmentation for obstruction learning. In some embodiments, a method comprises: obtaining images of obstacles that obstruct an object of interest; randomly scaling the images; extracting portions of the randomly scaled images at random positions in the randomly scaled images; and training a machine learning model using the extracted portions. In some embodiments, a method comprises: capturing images of obstacles that are at least partially obstructing an object of interest (e.g., a traffic light); processing the images using a machine learning model that is trained on a dataset that includes selected portions of augmented images of obstacles (e.g., large vehicles) that obstruct the object of interest; predicting a state of the object of interest based on output of the machine learning model; planning a trajectory for a vehicle; and causing the vehicle to travel the trajectory.