Research On Pedestrian Detection Algorithm Of Convolutional Neural Network Based On Enhanced Deep Features

Pedestrian detection is one of the most popular research topics in computer vision.This technology can be applied to AI robots,unmanned driving systems,video surveillance and other fields,and has great application value.Despite the rapid development of pedestrian detection technology,because the human body is non-rigid and is easily affected by conditions such as light and occlusion during detection,accurate detection of the person is still a major problem in this field.Recently,with the rapid development of deep learning technology,a large number of convolutional neural network pedestrian detection algorithms have emerged.This type of method extracts pedestrian features by using convolutional neural networks,and designs corresponding algorithm networks for this feature.Not only has the algorithm accuracy been greatly improved,it can also obtain excellent detection effect in scenes with dense crowds and poor lighting conditions.Although the convolutional neural network pedestrian detection algorithm has achieved great improvement in accuracy,but their models are complex and bring a large number of parameter calculations,which affects the detection speed of the algorithm and is contrary to the real-time nature of the pedestrian detection algorithm.Therefore,how to improve the imbalance of detection accuracy and speed in the pedestrian detection algorithm has become the most important problem faced by the current pedestrian detection algorithm.Under the one-stage framework of object detection,this dissertation uses a variety of feature enhancement methods to enhance the deep features of the network,and proposes a convolutional neural network pedestrian detection algorithm based on enhanced deep features,with the purpose of improving the imbalance between detection accuracy and speed.The main research work done in this dissertation and the results obtained are as follows.(1)The proposal of the algorithm which based on enhanced deep features.Firstly,by reading a large number of studies related to pedestrian detection algorithms,this dissertation summarizes the current research hotspots of pedestrian detection algorithms and the difficulties of pedestrian detection.Then,this dissertation conducts an in-depth analysis of the literature and summarizes the problems that still need to be improved in the pedestrian detection algorithm.Finally,aiming to improve the imbalance of detection accuracy and speed in the pedestrian detection algorithm,this dissertation proposes our algorithm based on the classic pedestrian detection model.(2)The design of network structure.This dissertation uses ResNet-50 as the backbone,builds on the network,and designs deep feature enhancement module,multi-scale feature extraction module,attention network module,feature pyramid fusion module,which together form the main part of the network of the dissertation.By adopting multiple feature enhancement methods such as receptive field enhancement for deep features,feature extraction at different scales,associated information mining in spatial and channel dimensions,and pyramid feature fusion,the information of deep features is enriched.At the same time,a stacked detection head and an improved loss function are proposed to further improve the detection accuracy of the algorithm.In addition,the network in this dissertation adopts a one-stage object detection framework,and uses small convolution kernels instead of large convolution kernels.At the same time,it adopts a feature channel compression technology,which greatly reduces the amount of network parameters and improves the detection speed of the algorithm.(3)Algorithm verification.In order to verify the effectiveness of the enhanced deep feature algorithm,this dissertation conducted multiple experiments which including ablation experiments and comparative experiments on two pedestrian detection data sets,City Persons and Caltech.The experiment was carried out under the Linux operating system and the Keras deep learning framework.By comparing the algorithm proposed in this dissertation with classic pedestrian detection algorithms and recent mainstream advanced pedestrian detection algorithms,the enhanced deep feature algorithm has better results in terms of detection accuracy and speed.Among them,in terms of accuracy,the algorithm in this dissertation has a Reasonable miss rate of 11.95% and 7.10% in the City Persons and Caltech datasets,respectively.In terms of speed,the algorithm in this dissertation processes 6.67 and 16.67 frames per second on the City Persons and Caltech datasets.There are three points of innovation in the whole research.(1)Specialized deep feature enhancement methods.This dissertation proposes a feature enhancement method that strengthens deep features,which enriches the information of deep features and increases the expressive ability of the network.(2)The efficient operation of multiple modules.By designing deep feature enhancement module,multi-scale feature extraction module,attention network module,feature pyramid fusion module,and stacked classification regression module,feature enhancement and classification regression are performed on deep features to optimize the performance of the algorithm.(3)Flexible network structure.This dissertation designs each functional unit in a modular form and encapsulate it in different sub-functions,which is easy to migrate to the training of other networks.