Deep Learning Based Object Recognition For Indoor Video Surveillance

The identification of human body target is an important part in the application scenario of the unmanned store.From the user's identification when entering the store to the user's identification in the store,it needs to ensure the real-time and accuracy of the identification at the same time.First of all,compared with the traditional method of target detection and tracking,this thesis uses the idea of "knowing target classification at a glance" for reference,and puts forward the research idea of "knowing who is at a glance" based on the algorithm of YOLO and the algorithm of image feature recognition.Based on this research idea,the identity recognition model of indoor monitoring object is designed.Secondly,the VOC data set is relabeled,so that the labeled data set includes human head and human body orientation.YOLOv3 training model configuration is improved.The improved network model can detect the human body target in the video and output the human body orientation feature in real time.For character feature matching,the design idea of multi-directional human feature extraction and classification training is proposed.Its implementation method is to extract the front,side and back features of human head at the same time,and classify the extracted feature vectors for training,and generate the human feature database of three directions at the same time.LBP algorithm is used for feature extraction,SVM algorithm is used for classification training.The main reason why the traditional machine learning algorithm is adopted instead of the deep learning algorithm with higher recognition accuracy is that in the application scenario of the unmanned store,the data collection and dynamic classification training of members before entering the store need to ensure the instantaneity,complete the data collection and classification training in the second level,and at the same time ensure the accuracy of feature matching.Finally,the identification model of indoor monitoring objects is tested and evaluated by combining video human body detection with multi-directional feature extraction,classification training and feature matching.This thesis,by improving the target detection training model,some progress has been made in the output of human head and human orientation features and the use of multi-directional human feature extraction and dynamic training.The experimental results show that when there are five human body classifications in the feature database,the dynamic training time of new customers before entering the store is0.26 seconds,which meets the real-time requirements of classification training.When there are five customers in the store,the speed of human object detection and identity recognition of surveillance video is more than 25 frames per second,which meets the real-time requirements of scene object recognition.When the characters are not occluded from each other,the recognition accuracy of the front and side of human body is more than 85%,and the recognition accuracy of the back of human body is more than 60%,through single camera video acquisition,the model can detect the human target and identity recognition in the indoor scene in real time,and output the identity information of members,which basically meets the design goal of model accuracy.At the same time,the experiment also shows that when the overlap of human targets exceeds a certain value,there is a blind area that can not be recognized by a single camera.When the appearance characteristics of human targets(height,body shape,clothing)are very close,the accuracy of feature matching will be reduced.Further research and experiments are needed,including multi-camera and multi-view target detection,multi-feature fusion algorithm or target tracking algorithm and other research directions.