ML Classification Correction for Vehicle Obstacle Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vehicle assistance systems struggle to accurately detect and respond to unfamiliar obstacles, as machine learning classification models cannot update their decision boundaries after deployment, leading to potential safety risks.
Innovation Solution
A computer-implemented method and apparatus that corrects output data from a machine learning classification model by acquiring images of encountered objects, providing them to a first model for initial classification, and then to a second model for discrimination between in-distribution and out-of-distribution data, with operator verification and annotation-based updates to refine the second model's decision boundaries.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a machine learning classification model is trained on predetermined training classes during deployment, then the system can accurately classify familiar obstacles, but it cannot detect or respond appropriately to unfamiliar obstacles (out-of-distribution data)
Solution Approach 1:
The system segments the classification task into two distinct models: a first machine learning classification model trained on predetermined training classes for accurate classification of familiar obstacles, and a second machine learning classification model trained to distinguish between in-distribution and out-of-distribution data for detecting unfamiliar obstacles. This segmentation allows each model to specialize in its respective function, resolving the contradiction between classification accuracy and adaptability to unfamiliar obstacles.
Solution Approach 2:
The second machine learning classification model acts as an intermediary between the first model and the driver assistance system. It receives output data from the first model and determines whether the encountered obstacle is familiar (in-distribution) or unfamiliar (out-of-distribution). This intermediary layer enables the system to handle both familiar and unfamiliar obstacles appropriately, maintaining classification accuracy for known objects while gaining adaptability to unknown objects.
2Reliability
If the machine learning model's decision boundaries are fixed after training, then the model provides stable and reliable predictions, but it cannot update or adapt to new types of obstacles encountered during operation
Solution Approach 1:
The system implements dynamic adaptability by allowing the second machine learning classification model to update its decision boundaries during operation. When out-of-distribution data is detected, the system can collect additional data and retrain the second model to expand its understanding of what constitutes unfamiliar obstacles. This dynamic capability enables the system to maintain reliability through stable in-distribution classification while adapting to new obstacle types over time.
Solution Approach 2:
The system incorporates feedback mechanisms where the second model's detection of out-of-distribution data triggers data collection and model retraining processes. The feedback loop allows the system to learn from encountered unfamiliar obstacles and update the second model's decision boundaries, enabling continuous improvement of adaptability while maintaining the stability of the first model's core classification function.
3Device complexity
If the system uses a single machine learning model for both classification and out-of-distribution detection, then the system structure is simple, but the model cannot simultaneously maintain high accuracy for familiar classes and detect unfamiliar classes
Solution Approach 1:
The system segments the detection task into two specialized models rather than using a single general-purpose model. The first model focuses on accurate classification of familiar obstacles using predetermined training classes, while the second model specializes in distinguishing in-distribution from out-of-distribution data. This segmentation improves detection accuracy for both familiar and unfamiliar obstacles compared to a single model approach, while maintaining manageable system complexity through modular architecture.
Solution Approach 2:
The second machine learning classification model serves multiple functions: it acts as a gatekeeper to identify out-of-distribution data, triggers data collection processes, initiates model retraining workflows, and provides alerts to the driver assistance system. This multi-functionality allows the system to maintain relatively simple structure while achieving high detection accuracy through the specialized capabilities of the second model.
Data Source
Figure 1~2
AI summary
A computer-implemented method (20) for correcting a first output data from a first machine learning classification model trained to associate an encountered object with a training class among predetermined training classes, the method comprising the following steps: a) acquiring an image of the at least one encountered object; b) providing the image to the first machine learning classification model, wherein the first output data are representative of an area delimiting each encountered object. c) providing the first output data to a second machine learning classification model trained to provide a second output data corresponding to the first output data discriminated between a first dataset and a second data set; d) verifying by an operator that the first dataset corresponds to one of the predetermined training classes representative data; and e) if the operator determines at least one discrimination error after said verification, updating the second machine learning classification model.