Design And Implementation Of Pedestrian Detection And Pose Estimation Based On Deep Learning

With the continuous development of machine learning and computer technology,deep learning has brought major breakthroughs in the fields of target detection and image classification.Pedestrian detection and pose estimation are the main research directions in the field of computer vision.Pedestrian detection and attitude estimation based on deep learning have become one of the most important methods,and they have important applications in the fields of smart transportation,video surveillance,and human-computer interaction.Pedestrian detection is the premise of pedestrian posture estimation.Pedestrian posture estimation is the basis of pedestrian action recognition.This article has done the following two researches on pedestrian detection and posture estimation:In terms of pedestrian detection,pedestrian detection is based on the Faster RCNN,a classic model of target detection.The feature extraction network VGG16 of the original Faster RCNN model is not sufficient to extract features.Only the deep network is used for extraction,ignoring the key fine-grained features of pedestrians.To solve this problem,this paper uses multi-layer fusion technology to integrate the Conv3?3,Conv4?3 and Conv5?3 layers of the VGG16 network.,The extracted features are fused with low-level texture details and high-level contour shapes and other prominent features.At the same time,pedestrian scales vary in size,and small-scale pedestrians have low resolution and are difficult to detect.This article improves the RPN module and uses 1×1,3×3,and 5×5convolution kernels of different sizes to slide on the feature map to generate candidate regions,Improve the detection accuracy of the model.In terms of pedestrian pose estimation,pedestrians are a very important and special group,and speed is also very important while ensuring detection accuracy.In order to balance the detection speed and accuracy,this paper is based on the improved Open Pose model for pedestrian pose estimation.The lightweight network Mobile Net is used to replace the original feature extraction network,and the calculation amount is reduced to one-tenth of the original network.At the same time,the cavity convolution is introduced to expand the receptive field and improve the accuracy.In addition,optimize the original network,share the repeated convolutional layer in the initialization phase,and replace the original 7×7 large convolution kernel with 1×1,3×3,3×3 in the refinement phase,and introduce a hole volume It maintains the same receptive field as the original network.The experimental results show that the amount of calculation is greatly reduced without loss of accuracy,and the speed of pedestrian attitude estimation is improved.