Occupancy Grid Fusion Using Machine Learning for Dynamic Cell Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous and semi-autonomous vehicle systems face challenges in accurately determining occupancy grids, which are essential for navigation and safety, as they often lose information from camera images and struggle to predict probabilities of dynamic occupancy grid cells effectively.
Innovation Solution
The system employs machine learning to determine a predicted occupancy grid based on previous grids and updates it using observed sensor measurements from multiple sensors, such as radar and cameras, to create an accurate and reliable occupancy grid, ensuring that probabilities of occupier types in each cell are accurately predicted.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional occupancy grid determination methods are used, then the system is simpler to implement, but information from camera images is lost and prediction accuracy of dynamic cells deteriorates
Solution Approach 1:
The occupancy grid is segmented into multiple cell types (free space, static objects, dynamic objects) with different probability representations. Each cell type is processed differently in the fusion algorithm, allowing preserved of specific information from camera images for dynamic cells while maintaining system feasibility.
Solution Approach 2:
The system implements feedback loops where the predicted occupancy grid is continuously updated with new sensor measurements. The machine learning model uses historical occupancy data to refine predictions, and the fusion algorithm iteratively adjusts probabilities based on sensor feedback, improving accuracy without complete information loss.
2Reliability
If machine learning is used to predict occupancy probabilities, then prediction accuracy of dynamic cells improves, but system complexity increases
Solution Approach 1:
The machine learning model serves multiple functions: it predicts occupancy probabilities, classifies object types, and provides uncertainty estimates. This multi-functionality reduces the need for separate specialized systems, managing complexity while improving prediction reliability across different cell types.
Solution Approach 2:
The system performs preliminary classification and probability estimation using the machine learning model before final occupancy grid determination. This preliminary action filters and prepares data in advance, making the subsequent fusion process more efficient and manageable despite the added model complexity.
3Reliability
If sensor data from multiple sources is integrated, then occupancy grid reliability improves, but processing complexity and computational load increase
Solution Approach 1:
The fusion algorithm applies different processing strategies to different regions of the occupancy grid based on local characteristics. Free space cells use simpler probability updates, while dynamic object cells employ more sophisticated machine learning-based predictions. This local differentiation improves reliability where needed without uniformly increasing processing complexity across the entire grid.
Solution Approach 2:
The system dynamically adjusts processing parameters such as probability thresholds, fusion weights, and model complexity based on environmental conditions and sensor data quality. This allows the system to maintain high reliability when conditions permit while reducing computational load when resources are constrained or conditions are poor.
Data Source
AI summary
An occupancy grid determination method includes: determining a predicted occupancy grid based on a previous occupancy grid, the predicted occupancy grid comprising a plurality of first cells corresponding to sub-regions of a region, each of the plurality of first cells including a plurality of predicted indications of probability each indicative of a predicted probability of a respective possible type of occupier of the respective first cell; determining, using machine learning and based on the first sensor measurements, an observed occupancy grid comprising a plurality of second cells corresponding to the sub-regions of the region; and determining an updated occupancy grid based on the observed occupancy grid and the predicted occupancy grid.