Multi-source Sensor Transportation Mode Recognition

In recent years,people’s action scenario recognition has become a popular research topic.Recognition people’s transportation modes is critical to personal scenario identification,as well as traffic control and urban planning.Currently,most traffic pattern recognition systems adopt Global Position System,Geographic Information System,acceleration,or a hybrid scheme of these sensors.However,these methods have limitations in terms of accuracy,power consumption and robustness.Against this background,this paper proposed novel scheme based on Bayesian polling and deep learning respectively,and studied pattern recognition problem for vehicles and other transportations.Main contributions are listed as follows:The first,original data set is obtained with the various sensors on smartphones.Then,to prepare for mode set extraction,the original data will be preprocessed and the mode set relate to transportation mode recognition will be obtained.According to different type of sensors,the characteristics are extracted,mined and analyzed from different aspects such as statistics,time domain,and frequency domain.The second to satisfy the low-power consumption and heterogeneity equipment requirement,this paper proposed a novel algorithm based on Bayesian voting sub-sensor.The scheme proposed to use sub-sensor to collect data,and utilize the various low-power sensors on smartphone(i.e.,acceleration sensors,gyroscopes,geomagnetism,base stations,and barometric pressure sensors)to collect data,then relevant features is extract from the data and classified by the Adaboost classifier,finally the transportation mode is recognized.Due to the probably mistake of the classifier,Bayesian voting is adopted in the decision layer,different weights are set for the sensors and the final result is obtained according to the voting results.Simulation and test result shows that the proposed method effectively improve the recognition accuracy,and the accurate is above 95%.The third to realize the high accuracy requirement,this dissertation proposed atransportation recognition algorithm based on deep learning.Deep learning has been widely applied in the field of mode recognition,especially in image processing.However,there is little research for deep learning in non-image data processing.Against this background,this paper proposed to apply deep learning in non-image transportation mode recognition.Different network structures is devised for various sensors based on the developed Keras structure and effectively improve the recognition performance of each sensor.Then the final result is obtained by voting.Simulation results reveal that this approach can achieve a high recognition rate by 94.18%.