Research On Pedestrian Detection Algorithm Based On Deep Convolutional Neural Network

Pedestrian detection technology belongs to the research category of computer vision,and aims to accurately identify and locate pedestrians in images through algorithms.Self-driving cars,intelligent robots,and surveillance security applications are all inseparable from the support of pedestrian detection technology.With the continuous development of deep learning,pedestrian detection technology has also entered a stage of rapid development,completing the transformation from relying on traditional methods to relying on deep learning methods.But what needs to be seen is that there are still many very challenging problems in the field of pedestrian detection that have not been completely solved.Among the various problems,the detection of small-scale pedestrians and the detection of occlusion pedestrians are two prominent problems.Small-scale pedestrians in images often have low resolution,which limits the pedestrian detection network to effectively extract its feature information,while occlusion the situation will cause the lack of pedestrian feature information,and the pedestrian feature itself has the characteristics of non-rigid deformation,which easily leads to missed inspections.In order to solve these more prominent problems in pedestrian detection,this article mainly made the following efforts:(1)In order to solve the problem that the shallow network of SSD algorithm cannot fully extract the feature information of small-scale pedestrians,a pedestrian detection algorithm based on SSD-Res Ne Xt50 network is proposed.After discussing the effectiveness of Res NeXt50 as the basic network of SSD,Res Ne Xt50 The basic network to replace SSD is VGG16,replace SSD's base network VGG16 with Res Ne Xt50.At the same time,by adding fusion module and prediction module,etc.,the detection performance of the algorithm is comprehensively improved.Finally,experiments were conducted on the Caltech pedestrian data set.In comparison with other similar algorithms,the superior performance of the algorithm in this chapter in small-scale pedestrian detection was proved.(2)In view of the difficulties in occlusion pedestrian detection,based on the good performance of the Repulsion Loss to deal with pedestrianocclusion problems,a pedestrian detection algorithm that combines the Repulsion Loss and the priori box recommendation strategy is proposed.According to the detailed characteristics of pedestrians,by improving the selection method of the a priori frame and setting the Io U threshold in a targeted manner,the accuracy of the prediction frame positioning and the learning efficiency of the network are improved.Experiments show that these improvements can significantly improve the pedestrian detection algorithm.