Timestamp Health in Multi-Sensor Driving Data: Measuring Jitter, Drift and Offset on KITTI, and What They Imply for ROS 2 Recording Pipelines

Multi-sensor perception stacks associate camera frames, LiDAR sweeps and inertial databy timestamp. Each sensor stamps data from its own clock, through its own software path,and when a stamp is wrong nothing fails: the data is simply mis-associated. This reportmeasures three timing faults, offset, drift and jitter, on three KITTI raw drives (16–33 s) andon synthetic faults with known ground truth. On real data the camera and LiDAR stampsare tight (robust interval jitter 0.19 ms and 0.08 ms), but the OXTS inertial stream containsthree backward timestamp steps, ten repeated stamps and two gaps. A nearest-timestamppairing estimator recovers injected drift exactly when stamps are noise-free, but with 2–10 %dropped frames its error standard deviation grows from about 8 ppm to 219–446 ppm unlessmis-paired samples are gated; gating restores 8–10 ppm. In simulation, resolving a 2 ppmdrift needs roughly 300 s of data, far longer than the drives analysed, so the health reportmust be able to say “not resolvable at this session length”. Each uncorrected millisecond costs0.45 cm of position and 0.009◦ of attitude on this drive. A gyro-versus-visual yaw estimatorreports a consistent apparent camera-to-IMU lag of 64–78 ms across three drives, which isnot what the dataset’s hardware synchronisation would lead one to expect; it changed signwhen the camera intrinsics were corrected, so we treat it as an unverified estimator output.We discuss what these findings suggest for ROS 2 recording and fusion pipelines and outlinea measurement protocol for follow-up work.