Attitude Estimation Based On Keypoints Detection

Deep learning technology was first applied to image classification,automatically learning and extracting deep features,so there is no need to manually design and extract features.With the development of technology,it has also exerted great value in the field of object detection,that is,based on the classification,it can also locate where the object is located in the picture.Of course,the development of technology is by no means satisfied with this.After accurately positioning and classifying objects,the extraction of more fine-grained features of objects has become a hot and difficult point of technology.Object segmentation and keypoint detection are similar in that they extract the fine-grained features of the object itself,except that segmentation is to extract the boundary line between the object and the background.Keypoint detection is to extract some more important parts of the object.The keypoint detection task depends on the results of the target detection and its accuracy,and the results of the keypoint detection can pave the way for subsequent technologies,such as pose estimation based on keypoints and timing action detection.This paper is based on key points to make attitude estimation.The whole process includes four aspects: target detection algorithm,single-person keypoint detection algorithm,pose estimation based on keypoints,and migration of human keypoint detection to hand keypoint detection.The main research contents are as follows:1.The object detection algorithm based on deep learning is studied.The mainstream target detection algorithms are mainly divided into single-step method and two-step method.The single-step method is fast and the two-step method has high accuracy.Considering that the subsequent detection of key points is extremely dependent on the accuracy of target detection,a two-step detection method faster-rcnn with higher accuracy is selected.The detection network and convolution operator are improved to ensure more accurate detection results.2.The algorithm of key point detection based on target detection results is studied.The key point detection algorithms are divided into top-down top-down and bottom-up bottom-up methods.The top-down algorithm has high detection accuracy,and the bottom-up algorithm has fast detection speed.This article selects top based on target detection.-down algorithm.This paper studies the accuracy and improvement of keypoint detection under the combined action of keypoint probability map heatmap and keypoint offset map.3.The application and improvement of migrating human keypoint detection algorithm to hand keypoint detection algorithm are studied.In this paper,there are 17 human key points and 21 hand key points.During the migration process,in addition to changing the number of key point detections,the input data enhancement method must be changed to make the migration effect better.4.The method of attitude estimation based on key points is studied.There are two methods for attitude estimation based on key points.One is to use deep learning methods,that is,to build a small classification network,input the detected key points,and output attitude categories.This method can To ensure high classification accuracy;another method is to manually design classification rules,that is,to design unique joint features for each pose to achieve classification.This method does not require large amounts of data,but the generalization performance is extremely dependent on manual design feature.5.Summarize the content of this research topic,and explain the research and development trends that can be further developed based on this research.