Human intent recognition is important to the control of robotic prosthesis. In this paper, we propose a multi-level real-time on-board system to recognize continuous locomotion modes. A cascaded classification strategy is designed for the recognition of six steady locomotion modes and 10 transitions. On-board signals of the robotic prosthesis include two inertial measurement units and one load cell. Three transtibial amputees are recruited in the experiments. The prediction decision time of the real-time on-board cascaded classification system is about 3.3 ms, which is enough short compared with the sliding window increment 10 ms. It is easy to recognize the standing and ambulation in the first-level classification with a 99.86% accuracy by quadratic discriminant analysis (QDA) classifier. In the second-level classification, threshold method is adopted to divide one stride into swing and stance phases. In swing phase, five steady modes are recognized with a total accuracy of 96.40%. In stance phase, all these five steady modes are recognized with a total accuracy of 91.21%. The average recognition accuracy of the three subjects is 93.21% by QDA classifier. Besides, for transitions, the proposed system could recognize all transitions rightly. The designed system is feasible and effective to realize real-time on-board recognition of continuous locomotion modes, which is promising for the further control of the prosthesis.