Research On Object Detection Algorithm Based On Center Keypoint Selection

Object detection is a long-standing topic in the field of computer vision.It is developed from the image classification task and aims to detect predefined categories of objects.After experiencing rapid development,it is roughly divided into two types: topdown object detection methods and bottom-up object detection methods.Among them,top-down methods treat each object as a prior point or a predefined anchor box,and then predict the relative offset of the bounding box.Most of this method uses the anchor box mechanism to improve the detection effect,so it will bring about the problems of many hyperparameters and imbalance of positive and negative samples.In recent years,the bottom-up method has been revived,which abandons the widely used anchor box mechanism,models the object as different parts,and then groups the parts of the same object together through some post-processing methods.The research content of this thesis belongs to the bottom-up object detection method,which is researched and improved on the basis of the classic CornerNet algorithm to further improve the performance of the bottom-up method.The research contents of this thesis are as follows:(1)Since the CornerNet algorithm detects the object as the top-left corner and the bottom-right corner,and then groups the corners through post-processing such as associative embedding,which makes the algorithm sensitive to the boundary of the object and lacks information in the candidate region.For this problem,by introducing an additional center keypoint detection,the center keypoint and corner pairs can be geometrically connected,thereby adding positional information in the candidate region.This can effectively eliminate the large number of false positives caused by only corner pair detection.(2)Since the quality of the center keypoint directly affects the accuracy of the candidate box,in order to ensure the high quality of the center keypoint,an adaptive center keypoint selection method is proposed.The method can generate independent thresholds for each center point according to the statistical characteristics of the center keypoint,so as to ensure that the integrity of all center keypoints can be considered while meeting their particularity.On the MS COCO2017 dataset,by combining the above two methods,the improved algorithm achieves 1.4% and 2.1% AP improvement at singlescale testing and multi-scale testing compared with the original algorithm,respectively.(3)The CornerNet algorithm uses the classic HourglassNet for human pose estimation as the backbone network,which has good results on keypoint detection.However,since only single-level feature maps are used for detection and lack of rich scale information,it is not conducive to detection of multi-scale objects.In this regard,a dilated convolution module is proposed to make up for the lack of receptive field by introducing dilated convolution,so that the final output feature map of the backbone network can contain richer scale information.On the MS COCO2017 dataset,the proposed algorithm uses HourglassNet-102 as the backbone network,and achieves an AP of 42.0% at singlescale testing and 44.5% at multi-scale testing,respectively.Figure [27] Table [11] Reference [78]...