Offline AV Object Tracking to Expose Phantom Tracks and Safety Gaps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicle control systems face challenges in accurately tracking objects due to limited compute budget and latency constraints, leading to phantom tracks and delayed detection of moving objects, which can compromise safety and efficiency.
Innovation Solution
An offline tracking system is introduced that replays runtime data to identify safety gaps and incorrect environmental understandings, allowing for the design of test cases and training of machine learning models within specific compute budget constraints, improving tracking fidelity and reducing phantom tracks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If advanced tracking algorithms are used to improve tracking accuracy, then tracking fidelity improves, but compute budget is exceeded and latency constraints are violated
Solution Approach 1:
The system performs preliminary tracking computations during offline data replay and simulation phases, preparing optimized tracking parameters and pre-computed object trajectories. This allows the real-time online system to use pre-prepared data and simplified algorithms, achieving high tracking accuracy without exceeding real-time compute budget or latency constraints.
2Measurement precision
If advanced tracking algorithms are used to improve tracking accuracy, then tracking fidelity improves, but device complexity increases
Solution Approach 1:
The tracking system is segmented into distinct offline and online components. The offline component handles complex computations including advanced algorithm execution, data replay, and simulation during non-critical periods. The online component uses simplified algorithms with pre-computed parameters, reducing real-time complexity while maintaining overall tracking accuracy through the division of computational tasks.
3Productivity
If online tracking data alone is used, then real-time performance is maintained, but tracking accuracy deteriorates due to phantom tracks and delayed detection
Solution Approach 1:
The system implements a feedback mechanism where offline tracking results and simulation outcomes are used to validate and correct online tracking data. Phantom tracks detected in offline analysis provide feedback to refine online tracking parameters, while delayed object detections from offline simulation trigger re-evaluation of online detection thresholds, continuously improving tracking accuracy without sacrificing real-time performance.
Data Source
AI summary
Disclosed are systems, apparatuses, methods, and computer-readable media to autonomous driving vehicles and, in particular, for tracking objects in an environment that an autonomous vehicle (AV) is navigating. A method includes receiving environment data from at least one sensor in an AV control system mounted to the AV, the environment data including online tracking data that identifies at least one object within the environment data that is recorded at drive time; annotating the at least one object from the environment data that are incorrectly identified by the AV control system; executing an offline tracking engine to generate offline tracking data that tracks the objects over time in the environment data; and identifying safety gaps between the online tracking data and the offline tracking data.