Intent recognition systems using pattern recognition technology to control powered lower-limb prostheses are promising for seamlessly changing between locomotion modes - such as transitioning from level walking to stair ascent. These transitions can be accomplished by training an algorithm to recognize the patterns of mechanical and/or myoelectric signals an amputee generates during and between different locomotion modes. While low error rates can be achieved with this method, it typically requires a substantial amount of training data to be gathered. To alleviate this burden, this study investigated training a user-independent classifier from a pool of lower limb amputees performing level walking, ramps and stairs on a powered prosthesis and tested generalization of the classifier to a novel subject. The effect of using the amputee's EMG signals in combination with the mechanical sensors on the leg was also evaluated for this user-independent classifier. Generalization was poor to a novel subject - 48% overall recognition rate with EMG and 62% without (mechanical sensors only). However, an important system improvement could be made by including a few level walking trials of the novel subject (only a few minutes of data collection) in the training data, the overall recognition rate improved to 86% with EMG and 83% without.