Research On Deep Learning Based Scene Detection With Phones

With the rapid popularization of smart phones,scene detection with phones provides an important basis for environment-related upper-layer applications.Due to the limits of memory and power of phones,the detection system requires high precision while also considering the energy consumption of the phone.Therefore,it is urgent to find a lowenergy detection technology with high accuracy.In recent years,deep learning has been successfully applied in various fields,with the advantages of strong learning ability,good adaptability and high accuracy.Therefore,this thesis introduces deep learning method into scene detection technology with phones,builds deep learning scene detection models with phones,and uses the real dataset collected in each scene for training and testing.The specific work is as follows:Firstly,different from the current research,this thesis no longer uses indoor and outdoor scene classification,but uses indoor/outdoor/underground scene detection in the light of the booming development of urban underground rail transit.The overlap of the characteristics of several sensors in the three scenes increases the difficulty of detection and puts forward higher requirements on the feature learning ability of the models.Secondly,this thesis innovatively proposes to learn the temporal correlation of sensor data.Therefore,we adopt long and short-term memory networks(LSTM),build LSTM based scene detection models,and explore the optimization of the networks.In addition,this thesis analyzes the contribution of low-power sensors in smartphones to scene detection and provides an important basis for future research on scene detection with phones.Thirdly,in order to meet the requirements of phone scene detection technology for low energy consumption and universality,we only use low-power sensor data that does not involve user privacy and is available in high,medium and low-end smart phones.In order to verify the performance of detection models in practical applications,we collect phone sensor data for each scene at different locations,time,weather conditions,and motion states.Finally,a scene detection system with phones based on deep learning is built with Tensor Flow and Python.The detection system is trained and tested based on the collected real dataset,and it is compared with five traditional machine learning methods including K-Nearest Neighbor(KNN),Support Vector Machine(SVM),Decision tree(DT),Logistic Regression(LR)and Random Forest(RF).Experimental results show that our proposed deep learning based scene detection model performs best.