This paper tackles a challenging problem of inertial sensor-based recognition for similar walking action classes. We solve two remaining problems of existing methods in the case of walking actions: action signal segmentation and recognition of similar action classes. First, to robustly segment the walking action under drastic changes such as speed, intensity, or style, we rely on the likelihood of heel strike that is computed employing a scale-space technique. Second, to improve the classification performance with similar action classes, we incorporate the inter-class relationship. In experiments, the proposed algorithms were positively validated with 97 subjects and five similar walking action classes, namely walking on flat ground, up/down stairs, and up/down a slope.