How early can we tell? : Conformal prediction for per-point anomaly detection at a roundabout junction

Sammanfattning

Detecting anomalous driving behaviour at complex junctions is an important component of intelligent transport systems, but most existing approaches output binary flags without a calibrated guarantee on the rate of false alarms, a property that is essential for operational deployment and increasingly required by regulatory frameworks such as the EU AI Act. This thesis investigates how conformal prediction can be applied to anomaly detection at a junction, using approximately two million trajectory points recorded by infrastructure-mounted stereo cameras at the Slottskogsgatan site in Gothenburg, Sweden. A pipeline was developed that derives the conflict-zone polygon from the empirical position density, engineers 39 trajectory features, constructs literature-grounded rule-based anomaly labels (sustained hard braking, anomalous stops, path deviation, close-conflict events), and validates two complementary conformal detectors against the held-out test set. Variant A operates per-row on six engineered features with a k-nearest-neighbour distance as the nonconformity score; Variant B operates per-trajectory on the directed Hausdorff distance to within-cluster training trajectories. Both detectors empirically honoured the calibrated false-alarm guarantee across five significance levels (ε ∈ {0.005, 0.01, 0.02, 0.05, 0.10}), with a single marginal exceedance for the per-trajectory detector at the tightest level. Variant A reached 25% recall on hard-braking events at ε = 0.05; Variant B reached 32% recall on path-deviating trajectories with a median lead time of 5.4 seconds, longer than Variant A's 3.0 seconds, reflecting that path divergence often precedes the kinematic event. A bias audit identified a 3.39× disparity in detector flag rate between imputed and clean sensor rows, surfacing a measurement bias that Mondrian conformal prediction by imputation status would address. The thesis contributes a validated methodological framework for calibrated anomaly detection at junctions and discusses the resulting system in the context of GDPR, the EU AI Act, and broader algorithmic-governance literature.