One of the challenges in the design of large-scale multiuser MIMO-OFDM systems is developing low-complexity detection algorithms. To achieve this goal, we leverage message passing algorithms over the factor graph that represents the multiuser MIMO-OFDM systems and approximate the original discrete messages with continuous Gaussian messages through the use of the minimum Kullback-Leibler (KL) divergence criterion. Several signal processing techniques are then proposed to achieve near-optimal performance at low complexity. First, the principle of expectation propagation is employed to compute the approximate Gaussian messages, where the symbol belief is approximated by a Gaussian distribution and then the approximate message is calculated from the Gaussian approximate belief. In addition, the approximate symbol belief can be computed by the a posteriori probabilities fed back from channel decoders, which reduces the complexity dramatically. Second, the first-order approximation of the message is utilized to further simplify the message updating, leading to an algorithm that is equivalent to the AMP algorithm proposed by Donoho <etal xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"/> Finally, the message updating is further simplified using the central-limit theorem. Compared with the single tree search sphere decoder (STS-SD) and the iterative (turbo) minimum mean-square error based soft interference cancellation (MMSE-SIC) in the literature through extensive simulations, the proposed message passing algorithms can achieve a near-optimal performance while the complexity is decreased by tens of times for a 64 <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex Notation="TeX">$\times$</tex></formula> 64 MIMO system. In addition, it is shown that the proposed message passing algorithms exhibit desirable tradeoffs between performance and complexity for a low-dimensional MIMO system.