Research For Pedestrian Detection And Person Identification Based On Deep Learning

Nowadays,intelligent devices are mainly responsible for users interaction and data collection.The consumption of network bandwidth and server computing are increased as intelligent devices are deployed more and more.In this regard,this reaserch builds a person identification system.The system consists of two parts,one is pedestrian detection based on Android platform,the other is person identification system based on computer.There are two advantages of this mode:1.It only transfers detected image,which saves network bandwidth.2.Pedestrian detection is implemented in Android device,which can relieve stress of server and contribute to the decentralization of power.In the research of pedestrian detection based on Android platform,we evaluated and improved several deep learning object detection algorithms,such as,Faster RCNN,YOLO,and so on.Due to weaker computational capabilities,we used Neon and OpenMP to accelerate convolutional neural network and detetion speed attach 0.1 second per frame.Due to improving algorithm and effective implementation of convolutional neural network,our Android pedestrian detection system is accurate and fast.In person identification system,we extract features of person,then use features to identificate person.Compared with training a classifier,this method is more flexible.The model trained by this method can be used anywhere.We use convolutional neural network to extract features,and the features can be deemed as ID of person.We test algorithm accuracy based on CUHK01 dataset,compared with multiple algorithm,our method attached 30.9% of highest top-1 accuracy.We cluster person through IDs,and the result demonstrate that our convolutional neural network has a strong distinguishing ability.Integrating theory and application,we compared and improved several deep learning algorithm.It decreaced calculation of parameters and increased identification accuracy.We used Neon,OpenMP to accelerate algorithm,and the calculation speed was very fast in two Android devices which have different computational capabilities.In a word,this research makes it fairly theoretical and is full of application value.In this experiment,algorithms and platforms is multiple.