D-BRANN: A Distributed Bidirectional Recurrent Neural Network for Full-Body Joint Angle Estimation from Sparse Inertial Measurement Units

Wearable Inertial Measurement Units (IMUs) enable full-body motion tracking studies beyond instrumented laboratory environments, which tend to be complex and rigid. However, current IMU-based motion tracking algorithms require complex neural network models that limit online deployment on the wearable device—or devices—due to hardware constraints. This paper explores the concept of distributing the computational load of a full-body joint-angle estimation algorithm across multiple wearable devices. We present a Distributed Bidirectional Recurrent Neural Network (D-BRANN) scheme that decomposes a sparse-IMU full-body joint angle estimator into five anatomical branches. Each branch processes one local IMU reading and a shared hip/root reference across three incremental stages: leaf position, joint position, and rotation estimation. Then, the distributed inferences of all branches are merged through a residual fusion algorithm. Compared to centralized benchmarks, D-BRANN requires around 66% fewer parameters and floating-point operations (FLOPs) per prediction. Despite this reduction, D-BRANN achieves SIP errors (mean global rotation errors of the upper arms and thighs) and whole-body angular errors that are non-inferior to those of the centralized pipeline benchmark within a 2° margin, with a SIP error of 14.88° compared to 14.95°, while its SIP error under fixed-window online inference increases moderately (18.66° compared to 16.69°). These results show that sparse-IMU pose estimation can be partitioned into substantially smaller computational components while preserving comparable offline SIP accuracy and real-time processing throughput, supporting future deployment across distributed wearable devices.