Highly automated vehicles (HAVs) must be rigorously evaluated before they are deployed on public roads. An accelerated evaluation framework was proposed in the literature to test HAVs more efficiently. However, running such a test is challenging due to the fact that some of the generated test cases may be not feasible or realistic. This paper proposes an improved accelerated evaluation framework that combines importance sampling with reachability analysis, so that the feasibility of all test cases are guaranteed, and the risk levels of cases are controlled. The performance of the proposed framework is studied using the unprotected pedestrian crossing scenario. A total od 2689 pedestrian–vehicle interaction events are extracted from open-source video data, and a truncated Gaussian mixture model (TGMM) is developed to describe the pedestrian–vehicle interaction. Simulation results show that the proposed method achieves unbiased crash rate estimation in an accelerated fashion while achieving the aforementioned benefits for test case generation (feasible and at controlled risk level).