Analysis of Sensor Fusion

Baseline performance of EagerMOT without attack indicates strong performance under many metrics including Higher-Order Tracking Accuracy (HOTA).

Under attack on camera where significant noise is added to bounding box detections, EagerMOT performance is undeterred in still maintains the same level of performance. This highlights the reliance of sensor fusion on 3D information in constructing 3D tracks.

When 3D information from LiDAR sensing is corrupted with 0.5 m of noise on average, the tracking performance is significantly degraded. Tracking outcome metrics are at unacceptably low levels despite small levels of object perturbation. This suggests, LiDAR sensing is safety-critical and an important module for security in AVs.