Polygon Sequence Matching for Accurate Vehicle Log Label Association
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous vehicle systems face challenges in accurately matching and associating polygon labels across frames due to the imprecision of current matching algorithms, which operate locally and fail to consider the sequence structure of tracks, leading to false negatives and positives.
Innovation Solution
A two-step process is implemented for polygon sequence matching, first performing frame-level matching to identify match-frames based on spatial and temporal nearness, followed by sequence-level matching to compute match quality scores, thereby generating candidate pairs for display in an output interface.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If local matching algorithms are used for polygon label association, then the matching process is simple and fast, but the accuracy is low due to false negatives and positives
Solution Approach 1:
The matching process is segmented into two distinct levels: frame-level matching that identifies potential match-frames based on spatial and temporal nearness, and sequence-level matching that computes match quality scores by considering the entire time series. This segmentation allows the system to handle complexity systematically while improving accuracy.
Solution Approach 2:
The patent transitions from local matching (two-dimensional frame-by-frame comparison) to sequence-level matching (three-dimensional temporal sequence comparison). By adding the temporal sequence dimension, the system can consider the entire track history rather than isolated frames, thereby improving matching accuracy.
2Productivity
If frame-level matching is performed to identify match-frames, then the process is efficient and identifies potential matches, but false matches may occur without sequence context
Solution Approach 1:
Frame-level matching is performed as a preliminary action to efficiently identify potential match-frames before the more computationally intensive sequence-level matching. This preliminary filtering based on spatial and temporal nearness reduces the number of candidates that require full sequence analysis, maintaining efficiency while setting up for more reliable matching.
Solution Approach 2:
The sequence-level matching process provides feedback by computing match quality scores that reflect the consistency of matches across the entire time series. This feedback mechanism allows the system to reject false matches identified at the frame level when the broader sequence context shows inconsistency, thereby improving reliability.
3Measurement precision
If sequence-level matching is implemented to consider time series, then match accuracy improves, but computational complexity increases
Solution Approach 1:
The computational task is segmented into two phases: a fast frame-level filtering phase that identifies potential matches based on simple spatial and temporal criteria, and a more intensive sequence-level scoring phase that is applied only to the reduced set of candidate matches. This segmentation significantly reduces overall computational time while maintaining high accuracy.
Solution Approach 2:
The system performs partial sequence analysis by computing match quality scores only for candidate pairs that survived the frame-level filtering, rather than exhaustively analyzing all possible frame combinations. This partial action approach achieves sufficient accuracy without the full computational burden of complete sequence comparison.
Data Source
AI summary
Disclosed herein are system, method, and computer program product embodiments for polygon sequence matching for label association in autonomous vehicle logs. A logged polygon is identified in each logged frame of a logged track. For each logged frame in the logged track, a labeled frame is retrieved from a labeled track that has a timestamp within a predetermined time window of the timestamp of the logged polygon. A similarity score is generated by comparing the logged polygon and the labeled polygon in the labeled frame. The logged frame and the labeled frame are identified as a match-frame when the similarity score meets a first threshold. A candidate pair is generated for each match-frame comprising the logged track and the labeled track that includes the labeled frame identified in the match-frame. The candidate pair with a match quality score meeting a second threshold is displayed.