Research On Object Detection Algorithm Based On Center Keypoints

As an important branch and foundation of computer vision technology,object detection has important application value and research significance in video surveillance,unmanned driving and human-computer interaction.With the continuous development of the principle of deep learning,the object detection technology based on this principle has gradually aroused people's eager attention.However,in the current object detection technology,candidate anchors are usually used to map the object area.This method causes the redundancy of the candidate anchors and increases the complexity of the network calculation.In order to solve this problem,on the basis of comprehensive domestic and foreign literature on object detection,this paper has conducted in-depth research on key point-based object detection algorithms.The main research contents of this article are as follows:(1)The object detection method based on keypoints is analyzed.Since traditional object detectors can no longer meet people's requirements for detection accuracy,this paper introduces keypoint-based human pose estimation algorithms into object detection,which eliminates the disadvantages of using anchor boxes in the previous object detection work.(2)The structure of the hourglass network model is studied,and it is considered that it consumes a lot of time in the inference stage,which results in the network failing to achieve real-time performance during detection.This paper proposes a new type of Refine-hourglass Network structure.In order to reduce the calculation of network parameters,the Fire Module is introduced by changing the structure of the original hourglass network.By comparing the effect of adding modules on the efficiency in the experiment,the results show that the detection efficiency of the Refine-hourglass network is significantly improved.(3)In order to further improve the network's problem of balancing inference time and accuracy,this paper proposes a discriminant method based on central keypoint to generate object detection frames.Different from other keypoint object detection algorithms,the algorithm in this paper only uses the keypoint information of the object in the middle position,that is,the center keypoint.This paper believes that the central keypoint information can not only reflect the global information of the object but also reduce the number of keypoints to significantly improve the inference time and accuracy of the network during detection.Finally,an experimental verification was performed on the MS-COCO public dataset,and the results showed that the proposed algorithm had improved results compared with other algorithms based on keypoints.