Design And Implementation Of A Specific Dress Tracking System Based On Deep Learning

With the progress of society,public security is also facing challenges all the time.Therefore,the government has increased monitoring facilities in public places.However,monitoring equipment can only store information,but can not analyze and process video.Cameras installed in public places often need to be monitored manually.When there are some emergencies or accidents,they need to be checked one by one,which requires a lot of manpower resources.An intelligent monitoring system can bring convenience to the society.The ordinary tracking system usually tracks the target through the pre-set tracking box,while the specific tracking system based on clothing can track the target selectively according to the tracking information provided by the user.Two ways are set up to build a user-friendly tracing system.One is to directly use computer vision technology to identify the target,through the pre-set tracking conditions to screen out the target,and then to meet the conditions of the target tracking;The other analyzes user input tracking text,using natural language processing to obtain specific information,and then using computer vision technology to track it.This intelligent monitoring system has certain practical value.The main research work is as follows.1)An instance segmentation data augmentation method based on clothing is designed.According to the characteristics of the monitoring scenario,the data augmentation method of mixed samples is used.Modified the training strategy and corrected the problem of improper labeling of public data sets.Different neural network models,evaluation metrics,and backbone networks were used to verify the results.A data augmentation method suitable for dressing application is established.2)A target tracking method based on clothing multi-information recognition is designed.Using Mask R-CNN,we can identify the characteristics of target categories and contours,and obtain the color clustering value of contour information using a clustering algorithm.The SVM classifier is used to obtain the color categories,which realizes the multi-information recognition of clothing colors and categories.Then the filter tracking algorithm is used to complete the follow-up tracking.The Deepfashion2 dress instance segmentation dataset was modified as the dress color dataset in order to make the dress color dataset closer to the real-world situation.3)Text-based dress information acquisition.The effect of directly using the Chinese word segmentation tool Jieba is average,so the Chinese word segmentation datasets PKU and MSR are used,the traditional and deep-learning Chinese word segmentation methods are compared.A tracking text dataset is constructed to verify the result,and the keywords after word segmentation are extracted according to the set rules to complete the text information.4)Target tracking system based on multi-information recognition of clothing.The target tracking system is designed and implemented for users.The system realizes the design of the front page,the interaction of the front end and back end,and the design of back end logic.