Research On Object Detection Algorithm Of High Density Pedestrian Based On Deep Learning

The changing trend of urban subway passenger flow is of great significance for urban traffic management scheduling,resource allocation and public safety,and pedestrian detection technology is the key technology to realize the change of passenger flow.At present,pedestrian detection technology based on traditional computer vision algorithm combined with machine learning has been widely used,but due to the high dependence on feature extraction,there are still problems in such aspects as noise interference,target occlusion and detection speed.With the image directly as input,the deep learning theory shows that the deep convolutional network can effectively realize data dimensionality reduction by using iterative training.Compared with the traditional method,the manual extraction of feature links is avoided,thereby,the influence of the different target background on the feature expression of the data is greatly reduced,and the detection precision and robustness of the system are improved.Based on deep learning,this paper studies the pedestrian detection technology in the subway monitoring scene.Aiming at the problem of target occlusion and overlap in high-density pedestrian scenes,design a reasonable sample labeling method,complete the statistical analysis of the labeled samples,optimize the anchor boxes parameters of the proposed anchor,and realize the faster region of interested pooling algorithm.The specific work is as follows:(1)Aiming at the problem of universality of target for target detection,we collected real passenger image data through the subway surveillance camera and taked manual labeling training and testing data sets for analysis.According to the concentrated distribution of the head and shoulder feature areas and the length-width scale ratio can be clearly divided into one type,we have improved the anchor box parameters in the Region Proposal Network(RPN)to make it more suitable for pedestrian detection in special scenes of the subway.The improved model has led to significant optimizations for missed and misdetected situations.(2)Aiming at the problem of occlusion of high-density pedestrians,a set of methods for labeling samples was designed by us.We define a classification rule of different samples and label region depending on the degree of occlusion.We finally got a data set containing tens of thousands of image samples(40,000 head and shoulder samples).The data is cleaned for the 10,000-person pedestrian image dataset.At the same time,we use the missed detection to mine the difficult samples,so that we can efficiently learn the model with excellent generalization performance.(3)We propose a solution for optimizing the speed of the model——Dilated Position Sensitive ROI Pooling,and it is used to reconstruct the model.We select the appropriate dilated hyperparameters through experiments.The algorithm can compress the neural network convolution model to a certain extent,and improve its detection speed without affecting the detection accuracy of the neural network convolution model.We can accurately detect the head and shoulders of different postures in the subway scene after deploying the model on the high-performance computing platform.In summary,we can accurately and quickly detect the number of high-density people in different scenarios and perspectives.And the algorithm verification was completed in the actual subway environment in Beijing.