Implicit Occupancy Prediction With Feature-Map Offsets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous systems face challenges in accurately and timely predicting occupancy in real-world environments to avoid collisions, particularly due to the computational expense of object-free approaches and the trade-offs in object-based systems between precision and recall.
Innovation Solution
A method and system that utilize a decoder model with multilayer perceptrons to predict occupancy at query points within a geographic region by generating point attributes through a feature map, combining sensor and map data without explicitly identifying individual objects, thereby reducing computational overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If object-based motion forecasting methods are used to predict trajectories for each detected object, then prediction accuracy for individual objects is improved, but computational cost increases due to processing each object separately
Solution Approach 1:
The patent combines object-based and occupancy-based approaches by integrating object detection results with occupancy grid predictions. The system uses object detectors to identify objects of interest, then applies occupancy network predictions to forecast trajectories for these detected objects, merging the strengths of both methods to achieve accurate predictions while managing computational resources efficiently.
Solution Approach 2:
The system segments the prediction process into two stages: first using object detectors to identify and filter objects of interest with confidence thresholding, then applying occupancy network predictions specifically to these segmented objects. This segmentation allows the computationally intensive occupancy method to be applied only where needed, reducing overall computational cost while maintaining accuracy for critical objects.
2Adaptability or versatility
If object-free approaches predict occupancy probability for each cell in a spatial-temporal grid, then detection coverage and expressiveness are improved, but computational expense increases significantly
Solution Approach 1:
The patent implements partial occupancy prediction by applying the occupancy network only to regions containing detected objects rather than computing occupancy for the entire spatial-temporal grid. This partial action approach maintains the detection coverage and expressiveness benefits of object-free methods while significantly reducing computational expense by limiting predictions to relevant regions only.
3Measurement precision
If threshold per-object confidence scores are used in object detection, then precision is improved, but recall deteriorates due to low-probability objects being filtered out
Solution Approach 1:
The occupancy network serves as an intermediary that processes low-confidence object detections and provides additional predictive information. Objects with low confidence scores are not discarded but instead fed into the occupancy network, which generates trajectory predictions that can inform the motion planner about potential low-probability objects and futures, thereby recovering information that would otherwise be lost.
Data Source
AI summary
Implicit occupancy for autonomous systems include receiving a request for a point attribute at a query point matching a geographic location, obtaining a query point feature vector from a feature map. The feature map encodes a geographic region that includes the geographic location. A first set of multilayer perceptrons of a decoder model process the query point feature vector to generate offsets. Offset feature vectors are obtained from the feature map for the offsets. A second set of multilayer perceptrons of the decoder model process the offset feature vectors and the query point feature vector to generate the point attribute. The operations further includes responding to the request with the point attribute.