Research On Pedestrian Detection And Human Pose Estimation In Natural Scenes

Pedestrian detection and human pose estimation are the research directions in the field of computer vision.Pedestrian detection is to detect the area where the pedestrian is located by analyzing the pedestrian characteristics in images or videos.Human pose estimation is to locate the position of each joint point in the image.Human pose estimation is also the basis of human action recognition and analysis.In this paper,a method based on convolution neural network is used to study pedestrian detection and human pose estimation.The main contents of this paper are as follows:(1)The classical Faster R-CNN target detection algorithm is studied.This algorithm has reached a high level in the common target detection task.Because of the similarity between pedestrian detection and common target detection,this paper applies Faster RCNN algorithm to the pedestrian detection task,and tests it on the Caltech pedestrian data set and compares it with several classical pedestrian detection methods.Faster RCNN-based pedestrian detection method under the "Reasonable" index missed detection rate of 23.3%,large-scale pedestrian missed detection rate of 7.4%,medium and small-scale pedestrian missed detection rate of 62.2%and 96.1%,respectively.The experimental results are superior to some classical pedestrian detection methods.(2)Faster R-CNN-based pedestrian detection method is not ideal for smaller scale pedestrians.To solve this problem,a multi-layer feature fusion mechanism is proposed,which scales the low-level and high-level feature maps and fuses information.The low-level feature maps,which are very important for small target objects,are fully utilized,and the missed detection rate under Reasonable index is reduced to 21.6%.After that,the size of pedestrian candidate box and interest pooling area was modified to make it more suitable for pedestrian characteristics,and the missed detection rate under Reasonable index was reduced by about 20.5%.Finally,aiming at the problem of difficult samples in pedestrian detection,the multitask loss function is replaced by focus loss function,which enhances the weight of difficult samples in the process of model training.The missing detection rate under "Reasonable" index is about 17.4%.The missing detection rate of small and medium scale is also reduced,and the performance of the improved algorithm is improved.(3)The algorithm principle of convolution attitude machine is studied.The algorithm uses full convolution network to extract feature from input image,and obtains a series of hot spots of human joints,and uses square error loss function to train the whole network.Aiming at the problem that the model volume of convolution attitude machine is too large to be transplanted to mobile terminal,the network is improved by using the idea of lightweight network Mobil eNet.The volume of the model is reduced from 124Mb to 19Mb.After that,in the basic fusion residual prediction module of convolution attitude machine,the accuracy of the model reaches 83.1%.Compared with the original model,the improved algorithm has the accuracy of the model.It can be transplanted to the mobile end,but the volume is reduced greatly.