Pedestrian Correlation Feature Recognition Based On Depth Learning

With the advent of a smart society and safe China,surveillance cameras are deployed in various scenes such as parking lots,traffic lanes,and banks,which brings convenience to maintaining public order and protecting pedestrians' safety,but also brings challenges.In a large number of images,only recognizing pedestrians can no longer meet the current social requirements,how to effectively use pedestrian limited features and accurately identify pedestrian correlation features has been put on the agenda.At the same time,the rapid development of deep learning,especially in the field of machine vision,is the most widely used.It not only has higher accuracy than traditional target detection and recognition methods,but also has high computational speed and has been applied in real life.Therefore,in view of pedestrian and pedestrian correlation feature recognition problem,the solution based on deep learning is more academic,effective and valuable.In the task of pedestrian and pedestrian correlation feature recognition for deep learning,the extraction of the target candidate region is the most important.In this thesis,we studies the method of extracting candidate regions by convolution feature map,which in the accelerating the region convolution neural network method.According to the dataset of this topic,a new feature extraction network structure is established to extract the features of original images.Secondly,it improved its original region proposal net,adopted different scales and different sizes of windows to extract proposal regions,solved the multi-scale of the target area,realized multi-classification tasks of region proposal net through IOU automatic labeling technology,and achieved the purpose of separating pedestrian-correlated feature areas.In the task of target recognition,Regional Convolutional Neural Network technology has demonstrated its power color.In this thesis,the regional convolution neural network is researched,and according to the dataset of this subject,a pedestrian correlation feature identification network structure is proposed.The network is used to extract the features of the correlated feature regions screened by the pedestrian correlated feature detection network,and to realize the recognition task of pedestrian correlated features,which including the style and color of clothing and style and color of the backpack.Finally,this thesis summarizes a series of research work done on this topic and prospects the further research work.