Deep neural networks used in autonomous vehicle perception are fundamentally uninterpretable black boxes that often predict incorrect classes with catastrophic overconfidence. Runtime Out-of-Distribution (OOD) safety monitors evaluate intermediate feature embeddings, Mahalanobis distance metrics, and energy-based scores to detect when visual input diverges from training data, triggering safety envelopes before an accident occurs.
Standard softmax outputs suffer from miscalibration: an out-of-distribution object like a kangaroo hopping on a Swedish highway is forced into an existing category (e.g. 'pedestrian') with 98% confidence. Energy-based safety monitors calculate Helmholtz free energy over the raw logit outputs: `E(x; f) = -T * log(Sum(exp(f_i(x)/T)))`. In-distribution inputs yield large negative energy values, whereas unknown anomalies exhibit high energy, allowing the safety manager to instantly veto the perception output.