Research On Indoor Positioning Algorithm Based On Multi-sensor Information Fusion

With the rapid development of technology and the improvement of people's living standards,service applications based on location information have received more and more attention.After extensive research,it is now widely accepted that the tendency of this field is to obtain different data information from multiple low-cost sensors,and use fusion technology to achieve accurate and robust positioning results.Many scientific research results have proved that image-based positioning is very good in scenarios with abundant features,but it will produce large and unpredictable errors in the situations with external disturbances.In recent years,inertial measurement units have become popular because of their high performance price ratio and their ability to achieve autonomous positioning.Its shortcoming is that the error accumulation problem will occur when inertial sensors run for a long time.In view of the tendency of indoor positioning and the advantages and disadvantages of visual and inertial positioning,this paper mainly proposes the following three fusion positioning algorithms from the perspective of different fusion structures:Firstly,an algorithm for integrating existing visual and inertial positioning results using the improved Kalman filter is proposed.The algorithm utilizes the distributed fusion structure to realize the fusion of visual and inertial positioning by designing a three state Kalman filter model which includes position,velocity and difference of acceleration,and introducing the concept of feature matching threshold.It makes full use of the merit of inertial positioning: the localization results of inertial localization are very accurate in a short time.Experiments show that the accuracy of fusion positioning is better than that based only on inertial data or visual features.At the same time,it can also solve the interference problems of the visual approach effectively.Secondly,a weighted localization algorithm based on global feature and differential zero-speed correction is proposed.Traditional inertial and visual positioning methods are both improved.Traditional visual positioning adopts local feature representation,which leads to very large calculation in the feature matching phase.This paper obtains global features through clustering method to reduce the search space in the matching phase and improve time efficiency.Then a differential zero velocity correction method is proposed by combining the idea of zero speed correction with the difference value characteristics of acceleration measurements under stationary and motion state.It can control the error accumulation of inertial positioning.Finally,the principal component analysis approach is used to assign weights to the improved inertial and visual positioning results.The algorithm has obvious advantages in time efficiency,and the positioning accuracy can meet the needs of most location-based services.Thirdly,a fusion positioning algorithm combining extreme learning machine and Dempster-Shafer evidence theory is proposed.The algorithm adopts the centralized fusion method.The input features including visual and inertial information are established firstly.Then the model is trained by extreme learning machine.Finally,the trust distribution of the positioning results is completed by Dempster-Shafer evidence theory.In addition,angle judgments are also introduced to decrease the big localization errors of turning.Experimental results show that the proposed algorithm can run in real time and at the same time achieve good localization accuracy even in challenging scenarios.