Autonomous Vehicle Trajectory Caching for Perception Gaps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicle perception and prediction systems face issues such as limited range or field of view, occlusion, oversegmentation, poor estimation of object position and derivatives, and incorrect classification, leading to intermittent inaccuracies that can cause indecisiveness in motion planning.
Innovation Solution
A method involving a perception system to detect objects, a prediction system to generate and score candidate trajectories, and a motion planning system to execute a final trajectory, with a persisted prediction cache to retain and update only trajectories that affect the final path, ensuring relevant context is maintained across planning cycles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Stability of the object's composition
If hysteresis is applied in decision making to bias towards previous actions, then decision stability is improved, but the ability to assess maneuver utility and plan trajectories is worsened due to insufficient context
Solution Approach 1:
The system performs preliminary action by persisting predicted object trajectories in a cache before motion planning decisions are made. This allows the system to have relevant context ready in advance, enabling both stable decisions (through hysteresis) and informed trajectory planning (through accessible historical context).
Solution Approach 2:
The persisted prediction cache acts as an intermediary between perception/prediction systems and motion planning. It stores and manages historical trajectory information, allowing the motion planning system to access relevant context without being directly coupled to the perception and prediction systems, thus resolving the information loss while maintaining decision stability.
2Loss of information
If all predicted object trajectories are persisted in the cache, then context availability is improved, but computational overhead and cache memory usage are worsened
Solution Approach 1:
The system extracts only the necessary information from the set of all predicted object trajectories. By persisting only those trajectories that actually affect the final autonomous vehicle trajectory, the system removes unnecessary data from the cache, reducing computational overhead and memory usage while maintaining adequate context availability.
Solution Approach 2:
Different treatment is applied to different predicted trajectories based on their relevance. Trajectories that affect the final path are persisted with full detail, while those that do not affect the path are excluded. This local differentiation optimizes the balance between context availability and computational efficiency.
3Device complexity
If perception systems detect objects with limited range or field of view, then sensor cost and complexity are reduced, but detection reliability is worsened due to occlusion and limited coverage
Solution Approach 1:
The system performs preliminary prediction of object trajectories based on current perception data. By predicting where objects are likely to be or move to, the system compensates for current limitations in sensor range and field of view, maintaining detection reliability without requiring more complex sensor arrays.
Solution Approach 2:
The prediction system acts as an intermediary that bridges the gap between limited sensor coverage and the need for comprehensive environmental awareness. By predicting object trajectories and persisting this information, the system effectively extends the functional range and reliability of the perception system beyond the physical limitations of the sensors.
Data Source
AI summary
Systems and methods for operating an autonomous vehicle (AV) are provided. The method includes detecting one or more objects in an environment, predicting a first set of predicted object trajectories comprising one or more trajectories for each of the detected one or more objects, generating a plurality of candidate AV trajectories for the AV, scoring each of the candidate AV trajectories according to a cost function, using the scoring to select a final AV trajectory for execution, determining which of the predicted object trajectories affected the final AV trajectory and which did not do so, adding the predicted object trajectories that affected the final AV trajectory to a persisted prediction cache, excluding from the persisted prediction cache any predicted object trajectories that did not affect the final AV trajectory, and executing the final AV trajectory to cause the AV to move along the final AV trajectory.