Research On Real-Time Pedestrain Detection Algorithm Based On Deep Learning

In recent years,artificial intelligence technology has gradually penetrated into all aspects of life.People's life is changing day by day because of artificial intelligence.As the basic application of artificial intelligence,real-time pedestrian detection technology based on deep learning provides technical support for the development of security,automatic driving,intelligent robot and other fields,and has broad application prospects.Based on object detector SSD,the paper optimizes it and realizes the task of real-time pedestrian detection.The main work includes the following steps:Firstly,the paper analyzes and summarizes the application scene of pedestrian detection algorithm,summarizes the research status,the key and difficult problems in pedestrian detection algorithm.Secondly,the paper firstly gives a brief overview of the traditional pedestrian detection algorithm based on machine learning,and analyzes the principle of commonly used machine learning related algorithms.Later,the paper focuses on the pedestrian detection algorithm based on deep learning,and introduces the two-step detection network and single-step detection network in the deep learning detection algorithm.Finally,the SSD network in the single-step detection algorithm is analyzed in detail,and the pedestrian detection data set is used to design the experiment as the benchmark experiment of the optimization algorithm comparison.The basic network structure of SSD is designed and optimized.Using the optimized residual module to build the basic network instead of the vgg16 network structure in SSD,not only reduces the calculation of the network,but also increases the depth of the network and improves the ability of feature extraction.Using the innovative feature fusion method to fuse the feature maps of different scales,not only more suitable feature maps are constructed,but also the number of feature layers used to extract the preselected box is reduced,and the real-time detection speed of the network is accelerated.At the end of the paper,the experimental verification of the proposed method is given.Finally,the problem of equalization of positive and negative samples in SSD algorithm is optimized.More difficult negative samples are obtained by adding negative sample data,and the ability of distinguishing difficult samples is improved.The loss function is optimized to make the selection of positive and negative samples more balanced.Different weights are used to calculate the loss value for different positive and negative samples.Finally,a new scoring function is proposed to rescore the prediction box to improve the detection rate of overlapping objects.At the end of the paper,the experimental verification of the proposed method is given.