Research On Pedistrain Accessory Detection And Retriveal Thechnology Via Deep Learning

In the intelligent video surveillance system,the detection and retrieval of pedestrian appendages is an important research topic in the field of computer vision.Pedestrian appendages are items attached to pedestrian such as bags,hats and so on.However,there are many challenges to this problem in the monitoring environment.On the one hand,there are many kinds of appendages with different shapes and large differences between individuals,which increases the difficulty of detection and identification.On the other hand,the changing light and perspective of monitoring scene making the appearance of the same object very different,which in turn poses a challenge for image-based appendages retrieval.In addition,in large-scale scenarios,the requirement of ensuring high retrieval accuracy and high retrieval speed at the same time is also a challenge.In this paper,we propose a pedestrian appendages detection model based on deep learning and a two-stage pedestrian appendages retrieval method.First of all,we use an end-to-end detection algorithm based on deep convolutional network to detect the appendages.To solve the problem that the small-scale target is difficult to detect,we design a candidate region extraction network based on Inception structure to generate more suitable proposals.The candidate region of the object is designed to classify and frame the regression using the coding network of local context information.Then,in order to meet the accuracy and speed requirements of appendices retrieval at the same time,we propose a two-stage pedestrian appendage s retrieval technology.In the first stage,we train a deep hash network which can get the deep hash of the image by changing the structure of the classification convolutional network.The deep hash feature extracted by the network is used to retrieve at a fast speed.In the second stage,we use the distance measurement network model to study the similarity function between images,and then extract the features of distance measurement network to calculate the similarity between images for accurate retrieval.Experiments show that our appendages detection method and retrieval method have achieved good results on our experimental data set.