Data quality in autonomous driving
Redundancy Multi-source and Multi-modal Data
Motivation
Autonomous vehicles depend on large streams of data from multiple sensors and modalities. Environmental conditions and sensor limitations affect data quality, but AV research has often emphasized algorithm design. My work asks how data quality should be measured in relation to the tasks and performance goals of an autonomous system, with redundancy as a concrete case study.
Approach
The Vase Framework connects five layers: data, data quality, tasks, applications, and goals. It maps quality requirements to task needs and system performance, opening questions about task orchestration and adaptive AV design.
The redundancy study uses nuScenes and Argoverse 2 to measure overlap in multisource camera data and multimodal camera–LiDAR data, then tests the effect of pruning redundant object labels with YOLOv8.
On three representative nuScenes overlap regions, selective label removal raises mAP50 from 0.66 to 0.70, 0.64 to 0.67, and 0.53 to 0.55. In Argoverse 2, removing 4.1–8.6% of labels keeps mAP50 near the 0.64 baseline. The analysis also identifies substantial camera–LiDAR redundancy. The findings establish redundancy as an actionable quality factor.
Papers & outputs
AV-1 — Multisource, multimodal autonomous-driving data: cameras, LiDAR, radar, vehicle telemetry, user data, and environmental context.
AV-2 — Research questions for measuring camera–camera and camera–LiDAR redundancy, curating redundancy-pruned datasets, and comparing detection performance.
AV-3 — Research designs for camera–camera and camera–LiDAR redundancy, showing pruning criteria and detection-model comparisons across redundancy levels.