Accurate identification of gait phases is a necessary step for control of robotic devices during gait therapy or automatic diagnosis of gait impairments. Most of the existing algorithms use a rule-based approach that takes advantage of the consistency of the gait cycle among healthy subjects. Since impaired gait patterns lack of that inter-subject consistency, most of those algorithms have limited performance when detecting phases in impaired subjects. In this paper, we propose a new algorithm for real-time detection of four gait events (heel-strike, toe-strike, heel-off and toe-off). The proposed algorithm uses a set of threshold-based rules and to compute the adequate values for the thresholds, maximizing the performance of the algorithm, we use a hybrid meta-heuristic approach that integrates Simulated Annealing and a Genetic Algorithm. Using data collected during overground and treadmill walking trials with a wearable device equipped with an inertial sensor, we report experimental results for three subjects: one healthy, one hemiparetic, and one myelopathic. F₁-scores for the three subjects were 0.98, 0.99, and 0.91, respectively.