Comparison of nonlinear attitude fusion filters

With the increasing percentage of embedded devices dedicated to analyse motion (robotic, IoT), it is crucial to assess which nonlinear fusion algorithm fits to a given application. We wonder if it is possible to design a framework for an extensive comparison of attitude fusion algorithms. In a simulated environment, we tested different patterns of motion, different types of noise and different algorithms, i.e: Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF), Quaternion Estimate (QUEST), Particles Filter (PF) and a nonlinear observer (CGO). We also developed a new variant of PF that was only theoretically examined. Because computing resources are often limited in an embedded context, we propose, here, two different scores, one more classical (s1) based on RMS attitude error and another one (s2) that is a compromise between accuracy and computation duration. We showed, on a simulated platform, that depending of the type of noise the s1 score can vary sharply. UKF and CGO are efficient for additive gaussian noise and PF is more efficient with different types of noise (multiplicative and impulsive). However, the s2 score is very stable with CGO that dominates the ranking and PF which is the last performer. We validated our results obtained by simulation with human motion thanks to real data from a smartphone device. In a context of additive gaussian noise for the sensors, we advice to use a non linear observer like CGO for embedded computation,for remote computation an UKF algorithm is a good choice and if the noise is not gaussian in a remote computation context the best choice is a PF algorithm.