Advanced powered lower-limb prosthetic devices require an intuitive and flexible user control interface to work in a dynamic environment. This study investigated the feasibility of inferring muscle activation patterns (electromyography, EMG, envelope) from non-invasive electroencephalography (EEG) signals. Six healthy individuals participated in this study; the subjects were instructed to walk at a comfortable speed across various terrains (e.g. level-ground, up/down slope, up/down stair walking). An unscented kalman filter (UKF) was used to predict the EMG envelope from fluctuations in the amplitude of slow cortical potentials of EEG in the delta band (0.1-3 Hz). The highest decoding accuracy obtained was an r-value (Pearson's correlation r-value) of 0.57 in the medial gastrocnemius of a single subject. In the same subject, the mean r-value across all the muscle groups exceeded 0.4. The mean accuracy across all subjects and muscle group corresponded to an r-value of 0.236. As for the Signal to Noise Ratio (SNR), 79.3% of the obtained results were more than 0 dB with mean performance SNR of 0.8 (max: 2.8 to min: -1.7). The highest accuracy was obtained using a lag of 50ms with a window length (tap) of 500ms. In conclusion, this is the first study to show offline continuous decoding of the EMG envelope during over-ground walking on multiple terrains. The results show the feasibility of such neural decoding. This method could be coupled with EMG-based terrain prediction techniques to further improve the neural control interface with powered lower-limb prostheses.