Abstract:

Due to the fast divergence of error in pedestrian track estimation， the PDR positioning trust time is short and the accumulated error is large. A pedestrian track estimation technology based on TrAdaBoost algorithm is proposed. In this method， outdoor pedestrian movement information is collected offline by GPS， and the TrAdaBoost algorithm in transfer learning is used to screen out the most suitable pedestrian movement features， which are transferred to indoor PDR location to correct pedestrian steps and realize pedestrian track calculation. The experimental results show that the pedestrians step roughly distribution within the scope of the 65—75 cm， and after correcting the positioning of the trajectory match degree is high， and real location will not be because of the accumulation of inertial sensor error and appear serious deviation， the average error compared with pure PDR localization algorithm has fallen dramatically， location accuracy is within 2 m of probability of 80%. Therefore， this method reduces the divergence speed of PDR， prolonging the credibility time of PDR location， and improving the credibility of the location results within a short distance.

Key words: migration learning, PDR localization, TrAdaBoost algorithm, step size correction