Generation of Target Gait Using Biomechanical Relational Network Based on Generative Adversarial Network (BMR-NetGAN)

Several methods have been developed to capture motion during real-time walking and provide feedback; however, these approaches may not always be suitable for every trainee owing to individual physical differences. In our previous study, we proposed a generative adversarial network (GAN)-based method to generate target gaits for active seniors. The generator incorporated the biomechanical relational network (BMR-Net) to extract inter-variable features of gait. However, the effectiveness of this block in generating individualized target gaits has not yet been verified. In this study, we examined whether the generator incorporating the BMR-Net (BMR-NetGAN) is effective in generating target gaits that reflect individual motion characteristics. In particular, we constructed a 2D transposed convolution GAN, which is generally effective for bidirectional feature extraction in the temporal and variable domains, and a generator without BMR-Net, and compared their results with those of BMR-NetGAN. The results demonstrated that BMR-NetGAN is an effective model for generating ideal gaits that reflect individual motion characteristics, as evidenced by adjustment patterns of lower-limb joint angles on the ZX plane that were not observed in the 2D transposed convolution GAN. Furthermore, an analysis of lower-limb joint motion indicated that BMR-NetGAN may successfully generate target gaits that account for left–right balance in individual participants.