Pedestrian Clothing Recognition Method Based On Deep Learning

Video surveillance is applied to all aspects of life and brings convenience to the government in maintaining social security.But there is also the problem that after a sudden accident,a large amount of surveillance video requires a lot of manpower to search for key video clips.With the rapid development of artificial intelligence,the technology applied to video surveillance is becoming more and more mature,making automatic identification and positioning possible.The pedestrians in the video are the focus of monitoring.Clothing is the most prominent feature of pedestrians.If effective clothing attribute recognition can be carried out,it is beneficial to the automatic identification and tracking of pedestrians.Not only can improve the efficiency of staff,but also important for the analysis of pedestrian behavior.Based on deep learning,this thesis conducts research on pedestrian clothing identification methods.The main work of this paper is as follows:(1)The research background and significance of pedestrian clothing attribute recognition are introduced,and the current development status of deep learning,pedestrian detection and clothing identification methods at home and abroad is summarized.The related knowledge of convolutional neural networks is expounded,including: macroscopic structure of unit layer and unit layer combination of convolutional neural network,loss function,regularization method,neural network parameter optimization method and 16-layer visual geometric group network model.(2)A multi-scale Pedestrian Detector(SSPD)algorithm is constructed to achieve high precision under real-time conditions.In the algorithm,the underlying feature layer is responsible for detecting a small proportion of pedestrians in the picture,and the high-level feature layer is responsible for detecting a larger proportion of pedestrians.This method helps to increase the recall rate of pedestrian detection,especially for smaller pedestrians.At the same time,a new aspect ratio prior frame(1:0.41)was designed to reduce the error rate and speed up the pedestrian detection.Then,by optimizing the original SSD loss function to better suit pedestrian detection,this not only eliminates the interference of the classifier but also reduces the time complexity.The simulation results show that the accuracy of pedestrian detection reaches 88.12 % when the frame rate reaches 20 fps.The comparison with other algorithms shows that the method is the best model for practical pedestrian detection,and it is an algorithm that can achieve balance in accuracy and speed.(3)Constructed a Multi-Task Stage Transfer Deep Learning(MTST)model,which can identify pedestrians under outdoor unconstrained conditions in the absence of outdoor pedestrian clothing data sets.Clothing attributes.The method uses a step-by-step learning strategy from human to learning,which is easy to difficult.The feature extraction of the source image with the tag annotation on the e-commerce website is firstly performed.The set loss function is useful for extracting more features of different attributes and then The learned attributes are migrated to the target image under unconstrained conditions.Combined with the pedestrian detection algorithm proposed in this paper,the foreground frame extracted from the pedestrian detection model is input into the model,and the comprehensive attribute recognition accuracy rate reaches 64.8 %.This method is compared with other typical clothing attribute recognition deep learning models,which is about 6 % higher than the currently accepted best FashionNet method in the case of less than 4000 pictures in the training set.The results show that the accuracy of MTST recognition of clothing attributes has been improved.