Maneuver Labeling for Autonomous Vehicle Sensor Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional prediction systems in autonomous vehicles rely on heuristic approaches, requiring extensive hand-tuning and are prone to errors with noisy sensor data and edge cases, and fail to accurately assign probabilistic confidence scores or utilize historical data from a fleet of vehicles.
Innovation Solution
A computing device assigns maneuver labels to sensor data from autonomous vehicles, generating a machine learning model that predicts object maneuvers by creating a weighted directed graph from candidate path plans, inferring intentions from turn signals and kinematic behavior, and using this model to operate the vehicle.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If heuristic approaches are used to generate predictions, then the system can operate with conventional methods, but extensive hand-tuning is required and prediction accuracy deteriorates with noisy sensor data and edge cases
Solution Approach 1:
The patent replaces heuristic approaches (mechanical/systematic rule-based methods) with a machine learning model that learns patterns from data. The system uses neural networks to process sensor data and generate predictions, substituting manual heuristic rules with automated learning-based approaches that adapt to noisy data and edge cases without requiring extensive hand-tuning.
Solution Approach 2:
The machine learning model performs self-learning from historical sensor data and manually labeled maneuver data. The system automatically improves its prediction capabilities through training on diverse driving scenarios, eliminating the need for continuous manual adjustment and hand-tuning of heuristic rules.
2Measurement precision
If conventional prediction systems are used, then the system structure remains simple, but the system fails to accurately assign probabilistic confidence scores to potential maneuvers
Solution Approach 1:
The patent replaces conventional deterministic prediction systems with a probabilistic machine learning model. The neural network outputs not just predicted maneuvers but also probabilistic confidence scores, enabling the system to quantify uncertainty and provide measured predictions rather than binary outcomes.
3Reliability
If conventional approaches are used, then historical sensor data from a fleet of vehicles is not utilized, but this would enable improved prediction through learned patterns from diverse driving scenarios
Solution Approach 1:
The patent merges data from multiple sources including historical sensor data from a fleet of vehicles with manually labeled maneuver data. The system combines these diverse datasets to train the machine learning model, leveraging collective experience from multiple vehicles to improve prediction robustness and generalization to edge cases.
Solution Approach 2:
The system performs preliminary training of the machine learning model using historical fleet data before deployment. By pre-training on extensive historical data from diverse driving scenarios, the model is prepared to handle various edge cases and noisy sensor data more effectively when deployed in production vehicles.
Data Source
AI summary
Various technologies described herein pertain to labeling sensor data generated by autonomous vehicles. A computing device identifies candidate path plans for an object in a driving environment of an autonomous vehicle based upon sensor data generated by sensor systems of the autonomous vehicle. The sensor data is indicative of positions of the object in the driving environment at sequential timesteps in a time period. Each candidate path plan is indicative of a possible maneuver being executed by the object during the time period. The computing device generates a weighted directed graph based upon the candidate path plans. The computing device determines a shortest path through the weighted directed graph. The computing device assigns a maneuver label to the sensor data based upon the shortest path, wherein the maneuver label is indicative of a maneuver that the object executes during the time period.