Research On Pedestrian Detection And Pedestrian Re-identification Based On Deep Learning

With the development of deep learning technology,pedestrian detection and reidentification technology is more and more widely applied to intelligent transportation,intelligent security,intelligent identification and other fields.Pedestrian detection and re-recognition of deep learning have made significant progress in accuracy and speed compared to traditional detection and re-identification methods.However,pedestrians detection and re-identification still face severe challenges due to adverse factors such as pedestrian occlusion,pedestrian scale being too small,and pedestrian perspective change.For the above problems,this paper has improved the pedestrian detection and pedestrian recognition methods.The main work is as follows.Pedestrian detection: For the general object detection algorithm SSD applied to pedestrian detection tasks,it is easy for small and medium-scale pedestrians to have missed detection problems.First,replace the SSD basic network structure VGG with a ResNet50 network with stronger feature extraction capabilities.Then the semantic information of the high-level convolution network is merged onto the low-level convolution network through the Top-Down Modulation structure.Experiments show that the optimization of the SSD network structure can improve the detection ability of small and medium-scale pedestrians.SSD is a universal object detection algorithm.in order to adapt to different scale object detection,SSD generates different width and height ratio default anchor in different convolutional layers.There are a lot of default anchor generated by SSD that do not match the pedestrian scale,making it difficult to optimize the network and it is difficult to converge.This paper uses a clustering algorithm to generate a default anchor on the pedestrian detection training data set.Experiments show that the default anchor is generated by clustering,which improves the matching degree between the default anchor and the pedestrian scale,reduces the number of default anchor,reduces the amount of calculation,makes the network prediction regression more accurate,and promotes the performance of the algorithm.Pedestrian re-identification: Firstly,we analyze and compare various pedestrian recognition loss functions,and perform feature visualization on Minst dataset.From the results of the distribution of the characteristics of the Triplet loss with batch hard mining,there are still a few sample features that deviate significantly from the center of this class.To this end,based on the idea of minimizing the distance intra-classes,the central loss constraint is added to Triplet loss with batch hard mining,and its intra-class constraints are enhanced.Experiments show that the improved Triplet loss with batch hard mining intra-class distribution is more compact.In order to realize the complementarity between the different loss functions,this paper discusses the synergistic effect of the cross entropy loss and the triple loss on the convolutional network.Contrastive experiments show that the combination of improved Triplet loss with batch hard mining and Angular softmax loss is more complementary and more coordinated,and the performance of the lifting algorithm is better.A simple pedestrian detection and re-identification system is realized,which can be used to locate the pedestrian activity area and has strong practical value.Cut 200 pictures in the film,make a data set,and test the performance of the system on the homemade data set.Experiments show that the detection accuracy and speed of the system can meet the business needs,but the recognition accuracy needs to be further improved.