Research And Application Of Pedestrian Detection And Re-identification Based On Convolutional Neural Networds

With the construction of smart cities and the popularization of public security,the amount of monitoring data is increasing.Artificial video surveillance can no longer meet people's needs.The application of intelligent monitoring technology has quickly become a research hotspot.Pedestrian detection and re-identification are hot research areas of intelligent monitoring.In recent years,with the development of deep learning,pedestrian detection and re-identification technologies have also improved.Based on this,this paper mainly proposes a probabilistic weighting algorithm based on coincidence frames for pedestrian detection problems,and makes improvements to suppress non-associated neurons for pedestrians' recognition problems and realizes the recognition of pedestrian behavior attributes based on multi-region full convolutional networks.The specific work is as follows:A Fast-RCNN pedestrian detection algorithm based on probability weighting of coincidence frames is studied and implemented.The pedestrian detection of the target image is achieved by two-level output.First,it is determined whether there is a pedestrian in the target image,and then the pedestrian's coordinates are calculated by the regression of the pedestrian image.In the process of coordinate calculation,the coincidence detection frame was used to increase the positioning accuracy of pedestrian coordinates in a probability-weighted manner,and experiments were conducted on INRIA and Caltech common data sets.The false detection rate of INRIA data sets increased from 14% to 8.3%,Caltech The false positive rate of the data set decreased from 12.86% to 11.96%.A pedestrian recognition algorithm based on non-relational neuron suppression was studied and implemented.In analyzing the process of recognizing pedestrians in convolutional neural networks,we found that not all learned features have positive and some even negative effects.In this paper,we propose a non-correlated neuron-reduced convolutional neural network for this chapter.Pedestrian re-recognition algorithm,the introduction of neuron correlation evaluation indicators,evaluation of the effectiveness of each neuron obtained from training,and inhibition of non-related neurons in the network training process.The experimental results on the common data set show that the convolutional neural network based on non-relational neuron suppression proposed in this chapter can improve the performance of pedestrian re-identification,especially for large sample size databases.A pedestrian recognition algorithm based on multi-region detection full convolutional network is designed and implemented.The network can extract multiple regions of interest and add multiple weights with similar characteristics to the target region to improve accuracy.The experimental results show that the algorithm can extract and weight multiple similarity feature regions during training,thus improving the recognition accuracy of pedestrian movements.In summary,this study proposes a fast-RCNN pedestrian detection algorithm based on coincidence-box probability weighting for pedestrian detection based on the deep learning framework of convolutional neural networks,and proposes unrecognized neuron suppression for pedestrian recognition based on detected pedestrian still images.The algorithm recognizes pedestrians,proposes a pedestrian motion recognition algorithm based on a multi-region detection full convolutional network,and performs experiments on different public data sets.The experimental results show that the proposed algorithm is effective and feasible for pedestrian detection and recognition problems.And improve the accuracy rate.