Multi-sensors Indoor Positioning Algorithm Based On Deep Learning

With the development of mobile Internet applications,the demand for indoor localization services keeps growing in the area of business,living,and health,for example,looking for a car in parking lots,shopping mall navigation,intelligent hospital guidance.However,the propagation of positioning signals in the complex indoor environment is affected by many factors such as occlusions,high pedestrian mobility,and so on.Precise positioning in complex environment has always been a hotspot in indoor localization.RSSI(Received Signal Strength Indicator)based fingerprinting and PDR(Pedestrian Dead Reckoning)are common approaches used in indoor localization.RSSI,describing coarse-grained wireless signal power information,is unstable and susceptible in complex indoor environment,which limits its positioning accuracy.Meanwhile,PDR utilizing inertial sensors such as accelerometer and gyroscope,also suffers from the drift error,which leads to the low accuracy of long-distance positioning.Therefore,RSSI fingerprint and inernial-sensor are usually combined to improve the positioning accuracy.Furthermore,the development of deep learning technology also provides a new method to extract accurate location features from multiple sensor information.The ResNet(RSSI/MAG)indoor positioning model takes RSSI and geomagnetic information as fingerprint,and adopts the residual network to establish the corresponding relationship model between fingerprint and location.The residual structure has the advantages of building deeper network,extracting deeper features and making model easier to train.To improve positioning accuracy,DeepMS(Deep Multiple Sensors),a hybrid localization model integrates CNN-LSTM and ResNet(RSSI/MAG)to fuse the information from the accelerometer,gyroscope,magnetic meter with RSSI.The data sequences generated by accelerometer and gyroscope are divided into non-overlapping windows,and convolution is adapted to extract features from the data of each window.LSTM(Long ShortTerm Memory)is utilized to capture time sequence relationship between different windows.The motion trajectory model of the positioning object is established.Then the final positioning output of DeepMS is the fusion of trajectory model result and fingerprint positioning result.Experiments based on actual test data show that compared with RSSI based ResNet positioning model and shallow CNN model,the deeper ResNet(RSSI/MAG)which adopts RSSI and geomagnetic information as signal fingerprint can improve the positioning accuracy.The DeepMS model,which introduces CNN-LSTM structure to integrate accelerometer and gyro information on the basis of ResNet(RSSI/MAG)can obtain the more precisely localization result.