Research On Keypoint Detection Algorithm Based On Deep Convolutional Neural Network

Keypoint detection is the basis of photoelectric imaging measurement and is a key technology for improving the tracking accuracy of the extended object and accurately estimating its movement posture and behavior intent.As an important area in Computer Vision,it is also widely used in the fields of behavior recognition,autonomous driving,VR,and AR.In recent years,research on human keypoint detection based on Deep Convolutional Neural Networks(DNN)is particularly hot,and a series of keypoint detection algorithms represented by HRNet have been developed.However,for the realtime and high-precision needs of exercise object pose measurement,these algorithms not only have to withstand the challenges of environmental light changes,target scale changes,viewing angle rotation,and background interference,but also have to solve the common problem of high computing complexity,lager model parameters,and lacking interpretability in DNN.Focusing on the development of high-precision,lightweight,and low-computing algorithms for motion object pose measurement technology.Based on DNN,this dissertation conducts the following research on keypoint detection:To meet the application needs of high-precision and low-computing networks,firstly,the redundant analysis and the de-redundant design are conducted on HRNet to remove the fourth stage which has a large number of parameters and high redundancy,forming a simple and direct de-redundant high-resolution network HRNet＿W32＿S3.To solve the performance degradation problem of the HRNet＿W32＿S3 due to the lack of 4th-scale features and feature fusion,a cascaded network named CHRNet＿W32 is then proposed,which is a concatenation of HRNet＿W32＿S3 and UNet module.The test results of the MPII dataset and COCO dataset show that with only 40% parameters of HRNet＿W32but getting the same accuracy as the former,CHRNet＿W32 not only removes most redundant structure but also greatly alleviates the performance degradation problem.To solve the performance degradation of the existing lightweight model,through the Mutually Enhancement Modeling m Ethod(MEME),the Efficient Baseline(EffBase)is reconstructed,and an Efficient and Effective Key Point Net(EEff KPNet)is innovatively proposed.Experiments on the MPII dataset and COCO dataset show that with only 14% parameters of HRNet＿W48 but getting a better accuracy than the former,EEff KPNet＿P0 not only achieves efficient and effective keypoint detection but also verifies MEME’s ability to solve the performance degradation problem for small models.To solve the problems of existing models that the lack of interpretability in the heatmap output and the insufficient robustness against data transformation,a TestingTime Augmentation method based on Aleatoric Uncertainty(TTA-AU)is proposed,which is a Plug-and-Play and can be used by various keypoint detection models,thus a series of enhanced keypoint detection algorithms using TTA-AU being constructed.A lot of experiments show that the enhanced algorithms can evaluate the uncertainty of the heatmap output and further improve the accuracy and robustness of keypoint detection.To apply and analyze the above networks or algorithms on the rigid body object,the keypoint detection dataset of fixed-wing airplane(Airplane)is constructed,which contains many representative difficulties and challenges.Then,the HRNet＿W32,CHRNet＿W32,Eff KPNet＿P0 networks and their TTA-AU enhancement algorithms are applied to this dataset for training and testing.Quantitative indicators and qualitative analysis show that each network and algorithm have obtained expected universal verification on fixed-wing airplane.Specifically,on the single card GTX-1080 Ti GPU platform,Eff KPNet＿P0 achieves efficient(22+ FPS for single-frame inference)and effective(87.3%AP and 90.0%AR)keypoint detection,and the TTA-AU enhanced algorithms obtain extra performance gain while still maintaining a fast reasoning speed.