Planning-Based Attention Signals for Autonomous Driving Scene Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles often process unnecessary sensor data, making them computationally inefficient as most sensor data are not relevant for the current driving scenario, leading to unused data that could enhance actor detection and localization accuracy.
Innovation Solution
A method is developed to generate planning-based attention signals by converting driving-scene data into a scene-graph and inputting it into a deep neural network to determine attention scores for actors in the scene, filtering out irrelevant actors, and adjusting scores based on time-to-collision and trajectory statistics, which are used to command the vehicle's trajectory.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the autonomous vehicle processes all sensor data, then the actor detection and localization accuracy is improved, but the computational efficiency deteriorates
Solution Approach 1:
The patent extracts and processes only the relevant sensor data that is actually needed for the current driving scenario. By using planning-based attention signals to identify which actors and sensor data are relevant to the current trajectory plan, the system extracts only the necessary portion of sensor data for processing, rather than processing all available sensor data. This resolves the contradiction by maintaining high detection accuracy for relevant actors while improving computational efficiency through selective data processing.
Solution Approach 2:
The patent applies different processing quality and attention levels to different regions and actors in the sensor data based on their relevance to the current driving scenario. Planning-based attention signals assign different weightings to different actors and sensor data points, with higher attention given to actors that are more relevant to the current trajectory. This local quality approach ensures high processing accuracy for critical actors while using reduced processing for less critical data, resolving the contradiction between overall accuracy and computational efficiency.
2Productivity
If the autonomous vehicle uses planning-based attention signals to filter sensor data, then the computational efficiency is improved, but the quantity of processed data is reduced
Solution Approach 1:
The patent implements a feedback mechanism where the planning algorithm generates attention signals based on the current trajectory plan, and these attention signals feed back to modulate which sensor data is processed. The attention signals provide feedback information about the relevance of different actors and sensor data points, enabling the system to dynamically adjust processing based on the current driving scenario. This feedback loop resolves the contradiction by ensuring that data reduction does not compromise the processing of critical information needed for safe autonomous driving.
Data Source
AI summary
A method generating planning-based attention signals includes receiving driving-scene data. The driving-scene data is indicative of a driving scene around a host vehicle. The driving-scene data includes map data and localization data. The driving scene includes a plurality of actors. The method further includes converting the driving-scene data into a scene-graph. The method further includes inputting the scene-graph into a deep neural network (DNN). Further, the method includes determining an attention score for each of the plurality of actors using the DNN and the scene-graph. The attention score of each of the plurality of actors represents a priority given to each of the plurality of actors in the driving scene. The method further includes commanding the host vehicle to autonomously drive according to a trajectory determined by taking into account the attention score of each of the plurality of actors.