Objective Two neural network algorithm models were constructed to estimate the three-dimensional (3D) ground reaction force (GRF) and center of pressure (COP) during walking, and the estimation results of the two algorithm models were compared, so as to provide a solution for the acquisition of gait dynamics data without force plate. Methods A total of 1 384 gait data were selected. Multi-layer perceptron (MLP) and convolutional neural network (CNN) were applid to construct models for estimating GRF and COP components based on the 3D trajectory of whole-body markers. 100 samples were randomly selected as the test set, and the estimation performance was evaluated by the correlation coefficient (r), relative root mean square error (rRMSE). Paired-sample t-tests were used to compare the estimation performance of the two neural network models. Results The r values of each components of GRF estimated by MLP were 0.954–0.993, and the rRMSEs were 4.36%–9.83%. The r values of each component of GRF estimated by CNN were 0.979–0.994, and the rRMSEs were 3.06%–6.69%. The r values of each component of COP estimated by MLP were 0.888–0.992, and the rRMSEs were 4.78%–16.63%. The r values of each component of COP estimated by CNN were 0.944–0.995, and rRMSEs were 3.06%–10.85%. The RMSEs of CNN in estimating the medio-lateral component of GRF , the medio-lateral and antero-posterior components of COP during right stance phase, as well as the medio-lateral and antero-posterior components of COP during left stance phase were all lower than those of MLP (P<0.01). The RMSEs of MLP in estimating the anterior-posterior component of GRF during right stance phase, as well as the anterior-posterior component of COP and the vertical direction of GRF during left stance phase were lower than those of CNN (P<0.01). Conclusions Both MLP and CNN achieved good estimation accuracy in estimating GRF and COP during walking based on the trajectory of whole-body markers. The estimation accuracy of MLP in estimating the anterior-posterior components and vertical component of GRF was better than that of CNN, while the estimation accuracy of CNN in estimating the medio-lateral component of GRF, the anterior-posterior and medio-lateral components of COP were better than that of MLP.