MiCaDe is a lightweight system that watches for drift in the relative pose between a LiDAR and a camera on an autonomous vehicle. It turns the problem into a statistical change‑detection task, operating on the geometry that would normally be used for 3‑D perception.

The method extracts several consistency metrics from matched 2‑D and 3‑D detections. These include horizontal and vertical 2‑D offsets (ex,ey), the overall magnitude of the mis‑projection (ea), the direction of systematic shift (esin, ecos), the overlap of the 3‑D box projected into 2‑D (ei), and the difference between depth from the 3‑D sensor and the camera (eb). For each short batch of frames the metrics are averaged, the variance of the directional components is calculated, and the resulting 9‑dimensional vector is scored against a Gaussian Mixture Model that represents normal, calibrated operation. The log‑likelihood of each batch is fed into a non‑parametric Change‑Point Model (CPM) that flags a persistent shift when the likelihood distribution changes.

On three public datasets—KITTI Raw, ZOD, and ONCE—and on a one‑hour real‑world run that was deliberately mis‑calibrated, MiCaDe detected mis‑calibration faster and with fewer false alarms than a supervised baseline from the literature. For example, on KITTI Raw with a target average run length of 10 000 seconds, MiCaDe’s false‑alarm rate dropped to 0.001 per hour while the detection delay stayed around 8.8 s, comparable to the 12 s delay of the baseline. It achieved similar performance on ZOD and ONCE.

The method relies on successful 2‑D–3‑D matching. When drift increases, the number of retained matches per frame falls from an average of five to three on ONCE, reducing the signal but not eliminating it.

MiCaDe’s false‑alarm control assumes that the batch‑level features are independent and identically distributed. In practice this assumption is only approximate, so the specified target run length is best viewed as a sensitivity knob rather than an exact guarantee.