Adaptive Environment Feature Detection for Fleet-Retrained AV Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of machine learning models in autonomous vehicles is their susceptibility to overfitting due to insufficient training data, leading to poor generalization and increased false positives/negatives, especially in challenging environmental conditions, requiring manual and time-consuming retraining that is inefficient and costly.
Innovation Solution
A method for continuous adaptive detection of environment features using a connected vehicle network to automatically retrieve and aggregate data from similar scenarios across multiple vehicles for incremental retraining of the machine learning model, enhancing its performance and precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If machine learning models are trained with limited training data, then training time and cost are reduced, but model precision and robustness deteriorate due to overfitting
Solution Approach 1:
The patent combines sensor data from multiple vehicles to create an aggregated training dataset. By merging data from different sources, the system achieves sufficient training data volume without requiring each individual vehicle to collect and process massive amounts of data separately, thus reducing training time while improving model precision through diverse training examples.
Solution Approach 2:
The server performs centralized data aggregation and model retraining that benefits all connected vehicles. A single retraining operation using aggregated data from multiple vehicles improves the model for the entire fleet, rather than requiring separate retraining for each vehicle, achieving multi-functionality and efficiency.
2Reliability
If machine learning models are retrained manually with sufficient diverse data, then model robustness improves, but time consumption and cost increase significantly
Solution Approach 1:
The system continuously collects and pre-processes sensor data from multiple vehicles in real-time, maintaining a ready pool of diverse training data. When model performance degradation is detected or corner cases are identified, the pre-prepared aggregated data can be immediately used for retraining, eliminating the time-consuming data collection and preparation phases.
Solution Approach 2:
The system implements feedback loops where model performance is continuously monitored, and when issues are detected (such as corner cases or performance degradation), the system automatically triggers retraining using pre-aggregated data from the server. This feedback mechanism ensures model robustness is maintained without requiring manual intervention or extensive retraining time.
3Measurement precision
If extensive manual labelling of training data is performed, then training data quality improves, but human effort and cost increase hugely
Solution Approach 1:
The system uses automatically collected sensor data from vehicles in normal operation as training data. The data is aggregated and processed by the server without requiring manual labelling or annotation. The vehicles themselves generate the training data through their regular sensing operations, eliminating the need for human labelling effort while maintaining high data quality through real-world operational diversity.
4Measurement precision
If machine learning models are trained with highly specific scenario data, then performance on known scenarios improves, but adaptability to new scenarios deteriorates
Solution Approach 1:
The server aggregates data from multiple vehicles that encounter diverse scenarios in their respective operating environments. This universal data pool includes various road conditions, weather conditions, traffic patterns, and geographic locations. The aggregated training data teaches the model to recognize multiple scenario types, improving both specific scenario detection accuracy and general adaptability to new scenarios.
Solution Approach 2:
The system transitions from single-vehicle, single-scenario training to multi-vehicle, multi-scenario training by adding the dimension of fleet-wide data aggregation. This dimensional expansion allows the model to learn from diverse scenario variations across different vehicles and environments, simultaneously improving specific scenario performance and overall adaptability.
Data Source
AI summary
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 method includes data collection by the ego-vehicle in which, based on the machine learning model of the ego-vehicle, the acquired sensor 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 are 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 own sensor data and process same, including images that are extracted, encoded and hashed.