Boundary-Aware Trajectory Prediction via Dynamic Model Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current techniques for determining object trajectories in autonomous vehicles face challenges in resource-constrained environments, particularly when objects are near boundaries or occluded, requiring efficient processing to ensure accurate predictions and safety.
Innovation Solution
The system employs a machine-learned model to classify objects based on sensor data, using a primary prediction component for objects remaining in the area and a secondary prediction component for objects exiting or becoming occluded, which reduces computational load by utilizing less data and contextual information accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a primary prediction component is used for all objects, then prediction accuracy is maintained, but computational load and processing time increase in resource-constrained environments
Solution Approach 1:
The prediction system is segmented into two distinct prediction components: a primary prediction component for objects remaining in the area and a secondary prediction component for objects exiting or becoming occluded. This segmentation allows the system to apply different levels of computational complexity based on the specific scenario, thereby maintaining accuracy where needed while improving overall processing efficiency.
Solution Approach 2:
The system dynamically selects between prediction components based on classification scores from the machine-learned model. When an object is classified as remaining in the area, the primary component is used; when classified as exiting or occluded, the secondary component is used. This dynamic adaptation optimizes computational resource usage while maintaining prediction reliability.
2Productivity
If a secondary prediction component is used for objects exiting or occluded, then processing speed improves, but prediction accuracy may deteriorate for boundary cases
Solution Approach 1:
Different prediction components with different levels of complexity are applied to different local cases: the primary component handles standard cases (objects remaining in area), while the secondary component handles specific boundary cases (objects exiting or occluded). This local quality approach ensures that each case receives the appropriate level of processing complexity, maintaining overall system reliability while improving efficiency.
Solution Approach 2:
The system changes the parameter of prediction model complexity based on the classification score. When the classification score indicates an object is exiting or becoming occluded, the system switches to the secondary prediction component which uses reduced data and contextual information, effectively changing the computational parameters to match the specific scenario requirements.
3Reliability
If full contextual information is used for all predictions, then prediction reliability is maintained, but energy consumption and computational resources increase
Solution Approach 1:
The secondary prediction component uses partial action by utilizing less contextual information compared to the primary component. For objects exiting or becoming occluded, the full contextual information is not necessary, so the system uses a reduced set of data that is sufficient for accurate prediction, thereby reducing computational energy consumption while maintaining adequate reliability.
Solution Approach 2:
The system dynamically changes the parameter of information completeness based on the prediction scenario. The machine-learned classification model determines whether full contextual information is necessary or if a reduced subset suffices, allowing the system to optimize energy consumption by processing only the necessary amount of data for each specific case.
Data Source
AI summary
Techniques for determining a classification score indicative of an object exiting an environment are discussed herein. Techniques may include analyzing sensor data associated with an area of an environment from a perspective, such as a top-down perspective. From this perspective, a machine-learned (ML) model may utilize techniques to determine output data including a classification score indicative of the object exiting the area and/or being occluded from a point of view of a vehicle within a threshold period of time. The classification score may be leveraged to determine whether a first prediction component or a second, less computationally expensive, prediction component may be employed to determine a predicted trajectory of the object.