In Wireless Body Area Networks (WBANs), on-body channels undergo temporal variations predominantly because of the body motion and shadowing of tissue. This paper focuses on a common human activity, walking, to explore the characteristics of the dynamic on-body propagation channel. A customized portable wireless sounder is built and utilized to collect the original on-body channel data both for the indoor and outdoor environments. We focus on exploring the second-order statistics of the channel gain variation during walking, including the temporal autocorrelation, the cross-correlation and the channel gain discrepancy between the two correlated links. This channel characterization facilitates more efficient designs and evaluations of WBAN systems, including more efficient network resource allocation and more efficient deployment of cooperative communication.