In this paper, we compute the residual energy of predictive frames (i.e., unidirectional and bidirectional motion prediction) by using the autocorrelation between successive video frames at different frame lags (or prediction terms). A new low-delay motion compensated temporal filtering (MCTF) structure is proposed based on the residual energy computations, which outperforms the `low-delay MCTF' scheme suggested in the Joint Scalable Video Model (JSVM) in terms of delay, number of badly matched blocks and PSNR performance. The PSNR performance of the proposed scheme is only slightly lower than the `conventional 5/3 MCTF' scheme. This scheme can also be used with other types of wavelet filters (e.g., Haar) and can substitute the MCTF in 3D wavelet-based video codecs to achieve low delay with only minor reduction in coding efficiency.