A Minimal-Input Framework for Cut-In Detection and Pair-Specific Risk Analysis in Highway Trajectory Data - Traj2Rel-SFC: Trajectory-to-Relation Reconstruction and Context Signatures for Cut-In Interactions

Sammanfattning

Highway cut-ins create new same-lane leader–follower interactions that may requirethe follower to brake, yet most analysis pipelines depend on dataset-provided lane andneighbour identifiers, limiting reuse across datasets. This thesis presents Traj2RelSFC, a minimal-input framework that reconstructs lane assignment and same-lanerelations from trajectory geometry (x,y) and lane-marking metadata, detects cut-insas explicit cutter–follower pairs, computes pair-specific surrogate safety measures(DHW, THW, TTC, DRAC), and encodes surrounding traffic context as reversible16-bit Hilbert space-filling-curve signatures.Evaluated on 60 highD recordings (with highD lane/neighbour identifiers usedonly as reference labels, and highD indicator columns used only once to calibrate afixed geometry convention for pairwise SSM computation), the framework achievesmean reconstruction accuracy >0.9998, cut-in detection F1 > 0.999, and contextsignature agreement of 99.49% over 1.4M stage rows. Multi-indicator analysisshows that THW, TTC, and DRAC capture complementary severity aspects; only0.08% of events exceed the hard-braking DRAC threshold. Decision-stage featurespredict execution-stage THW risk with ROC-AUC 0.82 under leave-one-recording-outcross-validation, and SFC context features alone achieve AUC 0.62, showing thatspatial context signatures carry standalone predictive signal, although they do notsignificantly improve AUC over the kinematic-only model. A small exploratoryextension on 10 exiD recordings further shows that the same lane-reconstructionand cut-in mining core can be moved to highly interactive highway entry/exit sceneswithout redesigning the pipeline, although this add-on is intentionally limited and isnot presented as a second benchmark.From a software-engineering perspective, the result is a portable and auditableanalysis pipeline: method inputs are explicitly restricted, intermediate relations arereconstructable and testable, and outputs can be reproduced without relying ondataset-specific derived fields.