Object Relevance Filtering for Low-Latency Vehicle Planning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In dense urban environments, autonomous vehicles face challenges in identifying relevant objects for operation planning due to high computational complexity and latency, as many detected objects are not directly related to vehicle operations, leading to inefficient use of computational bandwidth and potential safety risks.
Innovation Solution
A machine-learned model, known as a relevance filter, determines a relevance score for objects based on their impact on vehicle operations, filtering out irrelevant objects and reducing computational load by prioritizing only those that meet a threshold score, using a combination of convolutional and graph neural networks to process top-down environmental representations and object tracks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the autonomous vehicle tracks all detected objects in dense urban environments, then the object detection coverage is comprehensive, but the computational complexity and processing time increase significantly
Solution Approach 1:
The patent segments the set of all detected objects into two subsets: relevant objects that affect vehicle operation planning and irrelevant objects that do not. This segmentation is achieved through a relevance filter that processes object features and determines relevance scores, allowing the system to focus computational resources only on the relevant subset while maintaining comprehensive detection coverage.
Solution Approach 2:
The patent extracts and removes irrelevant objects from the tracking pipeline after detection. The relevance filter identifies objects that do not impact vehicle operation planning and excludes them from further processing, thereby reducing computational complexity without compromising the detection of important objects.
2Reliability
If the autonomous vehicle processes all detected objects for operation planning, then the decision-making completeness is high, but the processing latency increases
Solution Approach 1:
The patent applies preliminary action by determining object relevance before the operation planning process. The relevance filter pre-processes detected objects and identifies which ones will impact vehicle operation planning, so that only relevant objects are passed to the planning module. This preliminary filtering reduces the number of objects requiring detailed processing, thereby reducing latency while maintaining decision-making completeness.
3Measurement precision
If the autonomous vehicle uses human guidance to identify relevant objects, then the object relevance accuracy improves, but the operational autonomy decreases
Solution Approach 1:
The patent implements self-service by enabling the autonomous vehicle to automatically determine object relevance without human guidance. The relevance filter is trained on training data that includes sensor data, object tracks, and operation planning data, allowing the system to learn and autonomously identify relevant objects based on patterns in the training data, thus maintaining both accuracy and operational autonomy.
4Productivity
If the autonomous vehicle filters out irrelevant objects, then the computational bandwidth efficiency improves, but the risk of missing relevant objects increases
Solution Approach 1:
The patent employs feedback mechanisms where the relevance filter is trained using feedback from operation planning outcomes. The training data includes information about which objects actually impacted operation planning decisions, allowing the system to learn from past performance and continuously improve its relevance determination accuracy, thereby reducing the risk of missing relevant objects while maintaining computational efficiency.
Data Source
AI summary
A machine-learned architecture for determining whether an object is relevant to a vehicle's action planning may comprise a convolutional neural network, graph neural network, and/or multi-layer perceptron that may determine a relevance score associated with an object that indicates indicating whether an object is likely to impact operation(s) of a vehicle. In some examples, the machine-learned architecture may use scene information and/or an object track to determine the relevance score.