The Research And Implementation Of One-shot Learning In Object Detection And Visual Tracking

Object detection and visual tracking are two important tasks in the field of computer vision.They also widely used in life,such as object detection and tracking in intelligent driving and automatic monitoring of surveillance video.With the help of deep learning techniques,object detection task with large amounts of annotated data has achieved good results,but in some tasks such conditions can not be met,such as obtaining too many pictures of rare plants and animals.In this case,one-shot detection is still in the exploratory stage.The small amount of data,the diversity of individuals in the same category,the diversity of perspectives and environmental factors all pose great challenges to one-shot detection.Visual tracking and one-shot detection tasks are similar in that both of them have the commonness of one-shot learning,and they acquire features from objects in single pictures.But the task targets are different,the former focuses more on multiple objects of the same kind,while the latter focuses on the same object in different frames.The existing tracking framework mainly uses the mutual relationship to calculate the similarity of the objects on the inter-class,but has few comparison of characteristics in the objects on the intra-class.Aiming at the above problems,this paper studies one-shot learning,and combines the characteristics of object detection and visual tracking to carry out the framework design,effectively improving the performance of one-shot detection and single object tracking.First,given a sample picture,when there are only a few objects in the new picture that need to be detected(there is interference from other kinds of objects),a new framework is designed to identify and locate similar objects.According to the characteristics of traditional features and deep learning,object similarity is compared on contour andsemantic.Then according to the detected results we carry out optimizing in the second detection to get the final results.This method has achieved good results in Pascal VOC,UIUC CAR and Caltech 101 data sets.Second,when there are multiple objects in the picture that need to be detected and there are many disturbances,this paper proposes two-stage parallel comparison network(TSPCN)to speed up the detection speed and improve the accuracy of object detection.The feature pyramid is used to extract the proposals of different sizes(that is,the bounding box with possible target object),and the semantic features of different proposals are concatenated to return the similarity and border.In order to make better use of semantic information,the classification and regression methods are redesigned according to the particularity of one-shot detection task.The results of this method are improved on Pascal VOC,UIUC CAR and Caltech 101 data sets.The ability to process complex images was enhanced.Thirdly,by improving the existing visual tracking network structure and combining with the multi-layer semantic features,this paper compares the similarity of tracking objects from two directions,namely between classes and within classes.This method improves the tracking accuracy without losing the tracking efficiency.In summary,this paper proposes new one-shot detection frameworks and improves the existing visual tracking network by studying one-shot learning and combining the characteristics of object detection and visual tracking.Good results are achieved in both types of tasks.