Determining Spatial Relevancy of Objects for Improved Development of Multi-Object Tracking in Autonomous Driving Systems

The perception system for autonomous vehicles (AVs) typically outputs all the objects it can observe in a scene. This is significantly more objects than what the AV would interact with, and far more than any human driver focus on during a driving task. Validating the perception system on all the observed objects could penalize its performance based on objects that will never interact with the AV or affect its planned trajectory. This dissertation outlines a strategy for identifying a subset of objects, referred to as Spatially Relevant Objects (SRO), that the perception system must perform exceptionally well on. This is valuable for several reasons and has many applications. For example, it can be used to determine the set of objects that should be included in the verification dataset for the perception system and thereby have a more efficient development cycle. Additionally, when evaluating the perception system on the SRO subset, the computed metrics not only evaluate the performance but also consider the real-world safety of the perception system, and the results would heavily support the safety case arguments. Finally, determining spatially relevant objects using a representative dataset then plotting their footprints relative to the AV can help determine and confirm the necessary sensing requirements and field of view coverage by prioritizing areas where SROs are most likely to appear. This is done without ignoring the fact that the world contains objects of different classes with different kinematics and could behave in a non-compliant way. The preliminary finding of applying our system showed that we can measure the performance of the perception system using a subset that averages about 11 of the observed objects without compromising the safety of the AV.