Human Keypoint Detection Based On Deep Learning

In various image data in the field of computer vision,the parts related to people are usually the most important,resulting in various research hotspots such as face detection,pedestrian detection and person re-identification,and human keypoint detection.Among them,human keypoint detection has been widely used in intelligent surveillance,humancomputer interaction,standard motion analysis,etc.It is also a basic research subject in the field of computer vision,and research methods have changed from traditional methods to deep learning methods based on convolutional neural networks.This thesis focuses on the problem of multi-person keypoint detection based on deep learning.Using the top-down detection idea,it includes two important steps: pedestrian detection and human keypoint detection.They are studied and improved separately.The main research contents are summarized as follows:1.This thesis studies and analyzes the classical pedestrian detection methods.Based on Faster R-CNN(Faster Region-Based Convolutional Neural Network),it is improved from three perspectives that affect the pedestrian detection model.In network structure,more scientific residual network bottleneck structure is used and dilated convolution is introduced to increase the receptive field.In data processing,mixup that works well in the image classification field is introduced to augment the data to improve the generalization performance of the model.In loss function,aggregation loss is added to the original loss function to improve the crowded occlusion.It achieves high pedestrian detection precision in the COCO2017 dataset.2.This thesis deeply studies the human keypoint detection algorithms represented by Stacked Hourglass Network and Cascaded Pyramid Network,and analyzes some shortcomings of them.On this basis,starting from extracting more scales and more abundant features,a human keypoint detection network model with better performance is improved and obtained,which is called stacked dilated convolution pyramid network.The overall use of two-level network,reflecting the idea of coarse to fine detection.Skipping connections are widely used for feature fusion at different levels,so as to get more comprehensive features.3.For the human keypoint detection network used in this thesis,the processing method of loss function is improved to solve the human keypoints which are difficult to detect.Different processing methods are used for the output of different stage networks.The first stage network uses ordinary L2 loss to deal with all human keypoints,and the second stage network uses batch level hard keypoints mining method to deal with hard human keypoints.This thesis has carried out sufficient ablation experiments in the COCO2017 dataset for human keypoint detection task to verify the impact of different improvements.Finally,in order to make a more fair and reliable comparison,this thesis also submitted the prediction results on test-dev2017,and comprehensively compared with the state-of-theart results.The method of this thesis has achieved competitive results.