Research On Two-dimensional Object Pose Estimation Based On Key-point Detection

Two-dimensional pose estimation based on key-point detection is one of the most challenging tasks in computer vision.It aims to detect key-points of given objects in images,and obtain two-dimensional pose estimation results of the target according to the relations between the detected key-points.It has essential research value and widely application in behavior recognition,pose tracking and other aspects.In recent years,with the development of technologies related to deep learning,especially convolution neural network(CNN)as feature extraction models,the performance of methods of key-point detection and two-dimensional pose estimation has improved.However,there are problems about research objects and methods: firstly,the existing resources and achievements overfocus on human pose,while there are few research source and achievements on rigid objects such as fixed wing aircraft;secondly,although the performance of the CNN method is better than that of traditional method,the deepening of the network layer will leads to the increase of parameter quantity and operation time,which will impair the deployment of the model and the efficiency of two-dimensional pose estimation.This thesis mainly focuses on the following three aspects: the investigation and research on two-dimensional pose estimation methods based on keypoint detection,the lightweight modification of high-resolution feature extraction network,and the extension of the research object of two-dimensional pose estimation.Firstly,we investigate the methods of key-point detection and two-dimensional human pose estimation,review the methods of key-point feature representation and the operators and structures of CNN,analyze several typical two-dimensional human pose estimation methods by key-point representation and network architecture,and introduce related Data sets and evaluation indexs.Secondly,to aim at the problems of huge parameter quantity and low processing efficiency caused by high resolution maintaining and multi-stage cascade network structure in HRNet,which is the state of the art in two-dimensional human pose estimation.,an improved parallel HRNet structure is proposed.By analyzing the scale change of receptive field of HRNet-w32 by step,eliminating the redundant cascade stage in the structure,and referring to the U-Net structure with different up-sampling operators,the U-type parallel feature extraction module is designed.By analyzing the feature response of key-points,accuracy,processing speed and visualization results of the experiments of human key-point detection and human two-dimensional pose estimation,the best scheme of parallel HRNet maintains high accuracy of HRNet-w32,the parameter quantity is reduced to 38.5% of the original,and the processing speed is increased by 45.5% of the origin.It effectively improves the problem of HRNet,which is time-consuming feature extraction caused by excessive network parameters,and improves the comprehensive performance of the network structure.Thirdly,for given rigid object such as fixed wing aircrafts,of which research resources and method are limited in key-point detection and two-dimensional pose estimation.Referring to the annotation benchmark of human two-dimensional pose in COCO data set,the key-point annotation data set of fixed wing aircraft is created and the corresponding annotation benchmark is formulated.Based on the analysis of the formulation of two-dimensional pose estimation method for given rigid objects and the subsequent two-dimensional pose tracking implementation,this thesis proposes a scheme to realize two-dimensional pose tracking by fast detection of key-points through lightweight network,and designs a two-dimensional pose estimation method for rigid object based on the parallel HRNet structure.The performance of network training and two-dimensional pose estimation is tested on the fixed wing aircraft data set,and the two-dimensional pose tracking of fixed wing aircraft in video is further realized based on the pose estimation method.The results show that,compared with the original HRNet-w32,the improved HRNet has only 0.1% lower of accuracy,50.0% higher of processing speed,and basically obtains the task of two-dimensional pose tracking of fixed wing aircraft in the video,and achieves good results.Finally,we summarize this thesis and propect the future trending of twodimensional pose estimation based on key-point detection.