Adaptive Scenario Detection for Autonomous Driving Corner Cases
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models in autonomous vehicles face challenges such as overfitting and require extensive manual labor for training, leading to inefficient and costly retraining when encountering corner cases like tunnels or intersections, which affects their precision and robustness.
Innovation Solution
A method for continuously adaptive detection of environment features using a connected vehicle network where sensors data from similar scenarios are automatically collected and aggregated to incrementally retrain the machine learning model, leveraging image hashes and similarity scores to enhance model performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are trained with sufficient diverse examples to avoid overfitting, then model precision and robustness are improved, but training time and computational resources increase significantly
Solution Approach 1:
The system pre-processes and stores sensor data from multiple vehicles during normal operation, organizing it into structured formats with metadata. When a corner case is detected, the pre-organized data from similar scenarios can be immediately retrieved and used for retraining, eliminating the need for time-consuming data collection and processing during the retraining phase
Solution Approach 2:
The system dynamically adapts the retraining process by detecting corner cases in real-time operation and automatically initiating targeted retraining only when necessary. The training data selection is dynamic, focusing specifically on similar scenarios rather than retraining with all available data, thus reducing training time while maintaining model precision
2Measurement precision
If machine learning models are trained manually with extensive data labeling to improve accuracy, then detection precision is improved, but human effort and cost increase significantly
Solution Approach 1:
The system automatically detects corner cases during normal operation, retrieves relevant sensor data from the database, and initiates retraining without human intervention. The entire process from detection to model update is automated, eliminating the need for manual data labeling and reducing human effort while maintaining high detection precision through continuous adaptive retraining
3Adaptability or versatility
If machine learning models are retrained frequently to adapt to new scenarios, then model adaptability is improved, but computational resources and processing time increase
Solution Approach 1:
Instead of retraining the entire model frequently with all available data, the system performs partial retraining using only the specific sensor data from similar corner case scenarios. This selective approach provides sufficient adaptability for the detected corner cases while consuming significantly fewer computational resources compared to full model retraining
Data Source
Figure 1
Figure 2a
Figure 2b
AI summary
The invention refers to a method of continuously, adaptive detection of environment features in autonomous and assisted driving of an ego-vehicle connected to a server and to a plurality of other vehicles. The ego-vehicle and the other vehicles are provided with respective machine learning models that are similar between one another. The method comprises a data collection step by the ego-vehicle in which, based on the machine learning model of the ego-vehicle, the acquired sensors data - among which images, generate a specific scenario interpretation. When feedback is received that the scenario interpretation is wrong, the corresponding images of the ego-vehicle is encoded, hashed and sent to a server with the request to broadcast them to the plurality of other vehicles requesting the other vehicles to communicate the server specific similar scenario interpretations Then, the other vehicles identify the specific combination of environment features corresponding to the wrong scenario interpretation, acquire their owns sensors data and process same, including images that are extracted, encoded and hashed. The other vehicles compare the similarity between the image hashes received from the ego-vehicle and from each of the other vehicles Vi by means of a similarity score surpassing a given threshold, meaning that a high similarity score corresponds to similar data, in terms of structure and content. For the image hashes of the other vehicles where the given threshold is surpassed, the other vehicles identify a specific similar scenario interpretation and send it to the server. The server aggregates the sensors data of the other vehicles corresponding to the image hashes of the other vehicles and of the ego vehicle, incrementally retrains the machine learning model of the ego-vehicle and sends the incrementally - retrained model to the ego-vehicle together. The invention further refers to a trained machine learning model, to a system to carry out the method and to a plurality of non-transitory computer-readable mediums.