Pedestrian Detection Algorithm In Road Traffic Scene Based On Deep Learning

With the development of driverless technology and intelligent transportation,object detection has become the most typical problem in machine learning.Pedestrian detection,as an important part of intelligent transportation,has become one of the key research topics of researchers at home and abroad.This paper mainly focuses on pedestrian detection in road traffic scene,because the complex background and uncontrollable natural conditions will bring challenges to our detection task.Generally,the main performance is as follows: on the one hand,the scale and posture of pedestrians in the image are different,and at the same time,some pedestrians are affected by uncontrollable natural conditions such as external light and weather environment Fuzzy,especially makes the detection of fuzzy small-scale pedestrian more difficult;on the other hand,it will further increase the difficulty of later detection because of local occlusion,these problems and differences will affect the final detection effect.Whether pedestrians can be accurately detected in road traffic scenes depends on the performance of the core pedestrian detection algorithm.Therefore,based on these different problems,this paper designs different detection algorithm models based on deep learning method for research and verification:Firstly,in the road traffic scene with complex background,aiming at the problem of low pedestrian detection accuracy in general,in order to further improve the pedestrian detection accuracy,a two-stage algorithm fast r-cnn model based on region candidate is designed.By analyzing the detection effect and implementation of the model in the real scene,a deep residual network resnext model with stronger feature extraction ability is designed.Combined with the detection network,the overall model is built.Through the RPN generated feature map and candidate box combined with anchor box mechanism,it is transported to the classification regression layer for judgment.Compared with other models,its final detection effect is more advantageous.Then,a one-stage algorithm SSD model based on regression is designed to solve the problems of low accuracy and slow speed of pedestrian detection in road scene with fuzzy small targets and different poses.The resnext model is also used to replace vgg16 in the original SSD as the feature extraction network.Through the design of a priori box and two matching principles,and combined with the confidence score,the predicted targets are screened out.In addition,the non maximum suppression algorithm is used to remove and merge all the predicted rectangular boxes of the targets,so as to get the category and location of the targets.Finally,the detection results show that it is better than the original SSD Other detection algorithms of the same kind.Secondly,based on the previous two experimental results,an algorithm model based on deconvolution and multi-scale feature fusion is designed to study the problem of pedestrian being partially occluded.Firstly,densenet model is introduced and combined with resnext model as feature extraction network.Secondly,in order to make full use of context semantic information and obtain more details of the bottom layer,shallow low-resolution high semantic features and deep deconvolution module are fused by FPN fusion.For the output feature map,K-means + + clustering algorithm is used to do a clustering analysis for the features Finally,compared with other detection algorithms,the proposed algorithm achieves better results when pedestrians are partially occluded.Finally,based on GPU computing platform,the three models are verified by experiments.Two benchmark datasets INRIA and Kitti are used to verify the model qualitatively and quantitatively.At the same time,the model is compared with other classical pedestrian detection models.The experimental results show that the pedestrian detection model designed in this paper has good performance in accuracy and recall.