Safety evaluation of Autonomous Driving Systems (ADSes) requires identifying the conditions under which behaviour transitions from safe to unsafe. However, exhaustive exploration of the scenario parameter space is computationally prohibitive. To address this issue, we introduce a boundary exploration approach inspired by contour-tracing from image processing to efficiently reconstruct the collision boundary in two-dimensional scenario spaces. Our proposed approach operates in three phases: (i) an initialization phase that seeds the search either with randomly sampled scenarios or with scenarios sampled along the boundary of a reference driver model; (ii) a boundary-discovery phase that locates an initial collision/non-collision transition using either a breadth-first alternating search or a binary-search subdivision of the parameter space; and (iii) a boundary-tracing phase that applies Moore-Neighbour Tracing to contour the collision region while avoiding unnecessary sampling. We evaluate four variants of the approach over 35 search spaces of two logical scenarios, using Autoware as ADS, and compare them against exhaustive grid search. Overall, the approach identifies the safety boundary by sampling less than 11% of the search space, yielding nearly a tenfold speed-up with no loss of boundary fidelity. Moreover, it has a minimal overhead in terms of sampled scenarios that are not part of the boundary.