Research On Integrated Algorithm Of Human Detection And Appearance Attribute Recognition

In recent years,the amount of video data has exploded.How to obtain structured data from massive amounts of video data has become an urgent problem to be solved.Among the massive amounts of video data,people are most eager to obtain the location information and characteristics of pedestrians.Attribute information,such as pedestrian's gender,age,appearance and other characteristics,pedestrian attribute recognition as a basic computer vision task,plays an important role in image retrieval,pedestrian re-recognition,video data structuring and other fields.This article studies pedestrian attribute problems from two perspectives.First,in Chapter 3,we deeply research pedestrian attribute recognition based on the multi-label classification method,and then in Chapter 4,we study pedestrian attribute recognition based on the integrated method of pedestrian detection and attribute recognition.:In the direction of pedestrian attribute recognition based on the multi-label classification method,the improvement based on ResNet-50 network in this paper is as follows:1.In view of the problem that pedestrian attributes are distributed in different positions of pedestrian images,the attention mechanism module is introduced to help the model focus on the key areas of the image and improve the quality of the features extracted by the model.2.Aiming at the difficulty of identifying small target attributes,a feature pyramid fusion strategy is adopted to improve the recognition effect of the model on small target attributes.3.Aiming at the imbalance between positive and negative samples of pedestrian attributes,a loss function for multi-attribute recognition is proposed to improve the overall recognition effect of the model.The pedestrian attribute recognition network proposed in this paper was tested on the PETA dataset,with mA of 84.83%,accuracy of79.37%,precision of 87.47%,recall of 86.09%,and F1 of 86.77%.In the direction of pedestrian attribute recognition based on the integrated method of pedestrian detection and attribute recognition,this paper works as follows:1.Based on the universal target detection algorithm YOLO v3,the integrated network of pedestrian detection and attribute recognition is designed and implemented.2.Based on the integrated network proposed in this paper,three improvements are proposed:anchor settings,large-scale pedestrian data set pre-training,and the use of GioU loss function to improve the detection and recognition effect of the integrated network.3.Based on the integrated network proposed in this paper,design and implement real-time pedestrian detection and attribute recognition software,and demonstrate the running effect.After the pedestrian detection and attribute recognition integrated network proposed in this paper is improved,the inference time under the Nvidia Titan RTX graphics card is18ms,which is much faster than the cascaded attribute recognition network.It is tested on the WIDER Attribute data set andmAP₅₀ reaches 65.25%,mAP₇₅ Reaching 34.13%.