Research On Pedestrian Detection And Re-identification Based On Improved Convolutional Neural Network

Pedestrian detection and identification can determine whether a person enters an area and realize suspicious person tracking.It is very important for the development and application of intelligent security systems.With the advent of the fourth "industrial revolution",artificial intelligence represented by deep learning has developed rapidly,machine vision has achieved unprecedented breakthroughs,and has been applied in related fields such as intelligent monitoring and intelligent assisted driving.Among them,the pedestrian detection and re-identification technology in machine vision has received considerable research and attention.Based on the pedestrian image and the re-identification algorithm,this paper proposes a pedestrian detection and re-identification method based on improved convolutional neural network.Firstly,for the particularity of pedestrian detection,a single-shot multi-box detection algorithm is proposed for the detection of convolutional neural networks for pedestrian detection: one is to automatically select the aspect ratio of the default frame of the algorithm according to the data characteristics by using the K-Means algorithm.Convergence speed;the second is to increase the segmentation suggestion module,generate the identification tag of the target through the border label of the target detection,and solve the problem of insufficient pixel-level segmentation label;the third is to use tucker2 tensor decomposition and variational Bayesian matrix decomposition to train The subsequent convolution kernel is subjected to tensor decomposition to obtain a new convolution kernel module,which reduces the amount of calculation and increases the calculation speed while ensuring that the accuracy is not lost.Using the California Institute of Technology's pedestrian detection data set for training and test analysis,the effectiveness of the method was verified.Then,the pedestrian re-recognition algorithm is studied and a pedestrian re-recognition algorithm based on lock-relaxed singular value decomposition and reordering convolutional neural network is proposed.The algorithm has three improvements: one is to use the random erasure algorithm to the data in the network training is randomly processed to increase the training data,to improve the generics to a certain extent,and to reduce the possibility of over-fitting.The second is to adjust the parameters of the full concatenated layer of the traditional convolutional neural network.The problem is to use the method of lock-slack singular value decomposition;the third is to use a reordering algorithm to weight the experimental results,so that the results of the extracted features are further improved.The experimental results show that the method has achieved satisfactory results in pedestrian recognition.Finally,in order to study the repetitive calculation problem caused by pedestrian detection and pedestrian re-identification separately,a unified training test model for pedestrian detection and re-identification is proposed to realize pedestrian detection and reidentification of end-to-end feature sharing training detection.It can be seen from the experimental results that the model avoids some computational redundancy and improves the efficiency of detection and recognition.