The feasibility of decoding lower limb kinematics in human treadmill walking from noninvasive electroencephalography (EEG) has been demonstrated with linear Wiener filter. However, nonlinear relationship between neural activities and limb movements may challenge the linear decoders in real-time brain computer interface (BCI) applications. In this study, we propose a nonlinear neural decoder using an Unscented Kalman Filter (UKF) to infer lower limb joint angles from noninvasive scalp EEG signals during human treadmill walking. Our results demonstrate that lower limb joint angles during treadmill walking can be decoded from the fluctuations in the amplitude of slow cortical potentials in the delta band (0.1-3Hz). Overall, the average decoding accuracy were 0.43 ± 0.18 for Pearson's r value and 1.82 ± 3.07 for signal to noise ratio (SNR), and robust to ocular, muscle, or movement artifacts. Moreover, the signal preprocessing scheme and the design of UKF allow the implementation of the proposed EEG-based BCI for real-time applications. This has implications for the development of closed-loop EEG-based BCI systems for gait rehabilitation after stroke.