Autonomous Vehicle Prediction Models Using Modified Scene Map Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately predicting the behaviors of dynamic objects, especially in low-frequency scenarios or when encountering uncommon objects, due to limited training data, leading to reduced predictive performance and increased computational requirements.
Innovation Solution
The techniques involve modifying map data based on detected object features to generate a multi-channel representation, which is then input into machine learning prediction models, effectively replacing low-frequency objects with higher-frequency map features for which the models have been thoroughly trained, thereby improving predictive accuracy and reducing the need for extensive training data and resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning prediction models are trained on extensive data including low-frequency objects, then predictive performance improves, but training data requirements and computational resources increase
Solution Approach 1:
The patent creates synthetic training data by copying and transforming high-frequency map features to represent low-frequency objects. Instead of requiring actual training data for rare objects, the system generates synthetic examples by replicating and adapting data from common map features, thereby improving predictive performance without proportionally increasing training data volume
Solution Approach 2:
The system makes the prediction model universally applicable to both high-frequency and low-frequency objects by training on a unified dataset that includes synthetic representations of rare objects derived from common features. This multi-functional training approach allows the model to handle diverse object types without requiring separate extensive training datasets for each object frequency category
2Measurement precision
If machine learning prediction models are trained to handle low-frequency objects, then predictive accuracy for uncommon objects improves, but computational requirements increase
Solution Approach 1:
The patent reduces computational requirements by copying existing high-frequency training data patterns and adapting them to represent low-frequency objects. This synthetic data generation approach eliminates the need to collect, store, and process extensive actual training examples of rare objects, thereby improving predictive accuracy for uncommon objects without proportionally increasing computational resources
3Reliability
If map data is modified to include dynamic object features, then prediction model input quality improves, but data processing complexity increases
Solution Approach 1:
The patent merges dynamic object features from sensor data with static map features into a unified modified map data structure. This combination integrates information about low-frequency objects with the existing high-frequency map features, improving the overall input data quality for prediction models while using systematic integration processes to manage the complexity of data processing
Data Source
AI summary
An autonomous vehicle may modify scene context map data based on specific object types and/or features perceived within an environment, and may use the modified map data with prediction machine learning models to predict the behaviors of other dynamic objects in the environment. In some examples, the vehicle may receive sensor data of the environment, as well as map data representing the static context of the environment. The vehicle may analyze the sensor data to determine combinations of object types, features, and/or events, and may use predefined heuristics to determine modifications to existing map features based on the specific combinations of object data. A multi-channel representation of the environment based on the modified map data may be provided to prediction models to predict the behavior of dynamic objects and control the vehicle.