Road Feature Detection Using Artificial Shadow Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning models for detecting road features struggle with imbalanced datasets, leading to inaccurate detection of road features in shadowed regions due to insufficient training on a majority of well-lit images, resulting in high false negatives.
Innovation Solution
A system and method that generates artificial shadow data based on shadow characteristics, texture information, and random shapes to augment image data, training the ML model to accurately detect road features in both well-lit and shadowed regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the ML model is trained on a dataset with majority well-lit images and few shadowed region images, then the training data quantity is sufficient, but the detection accuracy in shadowed regions deteriorates
Solution Approach 1:
The patent creates artificial shadowed region images by copying well-lit images and applying shadow masks to generate synthetic shadowed versions. This allows the training dataset to be augmented with realistic shadowed region samples without requiring additional physical image captures, directly addressing the imbalance between well-lit and shadowed region training data.
Solution Approach 2:
The patent applies parameter changes by modifying image characteristics through shadow effects - specifically changing brightness, contrast, and color parameters in shadowed regions. By adjusting these visual parameters on existing images, the system generates diverse shadowed region training samples that maintain the original road feature information while presenting the challenging shadow conditions the model needs to learn.
2Ease of manufacture
If the ML model is trained on conventional balanced datasets, then the training process is simple, but the detection reliability in shadowed regions deteriorates
Solution Approach 1:
The patent performs preliminary action by pre-processing the training dataset to create shadowed region augmentations before the actual model training begins. By generating artificial shadowed images in advance and incorporating them into the training set, the system ensures the model is exposed to shadowed conditions during training without complicating the training process itself, thus improving reliability while maintaining ease of manufacture.
3Measurement precision
If more shadowed region images are collected for training, then the detection accuracy in shadowed regions improves, but the data collection complexity and time increase
Solution Approach 1:
Instead of physically collecting additional shadowed region images through field captures or aerial photography, the patent uses digital copying and transformation of existing well-lit images to create synthetic shadowed versions. This eliminates the time-consuming data collection process while still providing sufficient shadowed region samples for accurate model training.
Solution Approach 2:
The patent replaces the mechanical data collection process (physical image capture in shadowed conditions) with a computational image processing system. By using algorithms to generate artificial shadows on existing images, the system substitutes time-intensive field data collection with faster computer-based image synthesis, significantly reducing data collection time while maintaining detection accuracy.
Data Source
AI summary
An approach is provided for detecting road features. The approach, for example, involves receiving image data associated with one or more road features. The received image data may be associated with ground truth label data corresponding to the one or more road features. The approach further involves determining a set of shadowed regions from one or more images of the received image data. The approach further involves generating artificial shadow data based on the determined set of shadowed regions. The approach further involves augmenting the image data with the generated artificial shadow data by applying the generated artificial shadow data to the received image data. The approach further involves training a machine learning (ML) model based on the augmented image data and the ground truth label data. The ML model is trained to detect the one or more road features.