A Study On Child Pedestrian Behaviors Detection Based On Deep Learning

In recent years,with the development of my country’s economy and technology,transportation demand has also increased exponentially.The problems of road traffic congestion and frequent traffic accidents caused by the large number of small and medium-sized cars not only pose a huge challenge to the safety of public transportation,but also have a considerable negative impact on my country’s vigorous and sustainable economy.Statistics show that my country’s car ownership is only about 5% of the global car ownership,but the annual traffic accidents in our country far exceed this rate.The number of children who die in traffic accidents in our country will exceed 10,000 each year,and the death rate of child pedestrians in accidents is several times that of European and American countries.At present,there is insufficient attention and corresponding research directions for the high-risk group of child pedestrians in domestic and international traffic driving.This paper analyzes and compares a variety of target detection algorithm frameworks,chooses the darknet algorithm framework as the original deep learning basic network structure model under the open source YOLOv3 target recognition application framework based on deep learning.The main work of this paper includes three parts: database construction and pre-processing,child pedestrian behavior prediction system construction verification,and system performance optimization.The database construction process needs to collect and collect child pedestrian samples in different related databases,and supplementary shooting to collect school-age children’s picture data samples.Data pre-processing needs to select different image processing open source libraries for poor light intensity,blurred details,image angle tilt,The images with severe occlusion are screened and enhanced to improve the detection performance of the system.All the screened and processed samples are classified according to the six types of child pedestrian behaviors.The entire child pedestrian behavior prediction system is built and predicted on an adapted experimental platform.On the one hand,it analyzes the functional performance requirements of the system and the difficulty of technical realization,on the other hand,selects appropriate training parameters and training procedures to train the model,according to the industry The general measurement index compares and analyzes the test results of the trained model.Aiming at the shortcomings of the training model,the original data set is expanded by optimizing the relevant model training parameters,while using the hard-case expansion and miss-identification iterative expansion methods.By comparing and analyzing the detection results produced by different weight models,the best detection effect is selected.The model as the system model.The optimized model has an average accuracy rate of 74.17% for each category,and a detection speed of 27 fps,which meets the expectations of system functional performance requirements.