Researches On Pedestrian Detection And Human Keypoint Estimation Algorithms

With the rapid development of deep learning in computer vision and the improvement of people's safety awareness,video surveillance systems has become more extensive and intelligent.Video Abstract is an effective way to obtain relevant information from a large number of original surveillance videos,and pedestrian detection and human keypoint estimation techniques are the basis for implementing video abstract.Pedestrian detection and human keypoint estimation are the position information of the pedestrian or human key points in the image,which is one of the difficulties in the field of computer vision research.This paper investigates the traditional pedestrian detection and human key point estimation methods.The existing pedestrian detection algorithm is not ideal for the processing of occlusion or poor illumination images,and the detection result has a high miss detection rate.The traditional human keypoint estimation algorithm often ignores the problem that the scale information of different parts of the human body varies greatly with the change of the human body posture and the shooting angle.Therefore,this paper has done the following work.In the occlusion or poor illumination image,the pedestrian detection result in the surveillance video often has a large miss detection rate.Therefore,the pedestrian detection method based on the model fusion algorithm is designed and optimized.In this paper,the traditional pedestrian detection method based on HOG+SVM is researched and the generation strategy of candidate boxes in the detection process is optimized.Then the deep learning pedestrian detection method based on GoogLeNet network is researched and the fine-tuning strategy of migration learning is used to generate its own pedestrian detection model.Finally,the pedestrian detection method based on the model fusion algorithm is designed and optimized.At the same time,in the case of poor illumination or occlusion,this paper verifies that the method has better pedestrian detection performance through the open test set.Aiming at the scale difference problem of human keypoint estimation,this paper designs and optimizes the human keypoint estimation method based on FCHN network.Based on the existing FPN network,this paper designs and optimizes the FPN network based on the unpooling layer,which realizes the extraction of the feature maps that combine the shallow features and deep features in the image and accelerates the sampling speed through the unpooling.Hourglass Network based on deconvolution layer was designed and optimized,and the learning ability of the model was enhanced by increasing the network learning parameters.Finally,based on the advantages of the above two models,the keypoint estimation method based on FCHN network was designed and optimized.The experimental results show that the @0.5 value of the method on the MPII dataset is 0.6% higher than the original model,especially the result of the ankle detection is increased by 1.2%,which solves the scale difference problem to some extent.On the premise of existing technology,the pedestrian detection method based on the model fusion algorithm and the human keypoint estimation method based on FCHN network are integrated into the system,and the pedestrian detection and human keypoint estimation system is built.The performance of both methods and the system was tested by experiments.