Research On Algorithms Of Pedestrian Detection And Crowd Counting Based On Computer Vision

In the field of computer vision,there are many problems with pedestrian-oriented research.Pedestrian detection and crowd counting are two of the major research focuses.They have a wide range of applications,including smart driving and intelligent monitoring.This thesis conducts further research and improvement on the major issues existing in the field of pedestrian detection and crowd counting.The main contents are as follows:First,the algorithms of pedestrian detection and crowd counting researched in this thesis are all based on deep convolution neural network(CNN),so we firstly introduce the development history,basic structure,and typical network models of CNN to help algorithm research in the following sections.Second,we propose a pedestrian detection architecture based on convolution feature pyramid and Single Shot MultiBox Detector(SSD).To solve the problem of high miss rate of pedestrians in detection task,we use convolutional feature pyramid to fuse spatial information of low-level convolution features and semantic information of high-level convolution features,and we find that the convolution feature pyramid can significantly reduce the miss rate of pedestrians.In order to further improve the detection accuracy of small targets,we combine the low-level(non-detection layer)convolutional features with the detection layer's features through the cross-channel rearrangement of local features to reduce the miss rate of small pedestrians.In addition,the network performance is further improved by introducing a Dense Convolutional Network(DenseNet)that performs better than VGGNet.Third,we propose a crowd counting architecture based on VGGNet and MultiColumn Neural Network(MCNN).Aiming at the complicated training process of MCNN and the risk of over-fitting,this thesis proposes to use the transfer learning of pre-training network-VGGNet to simplify the network and the training process.The VGG-MCNN method can effectively improve the network's accuracy and reduce the risk of over-fitting.At the same time,in order to reduce the loss of accuracy caused by large-scale downsampling,we use transposed convolution to upsample the feature map,and we find that our method can significantly improve the accuracy of the crowd counting algorithm.Fourth,in order to improve the real-time performance of pedestrian detection and crowd counting,we use a model acceleration method based on depthwise separable convolution,and obtains significant results in speed improvement of pedestrian detection network and crowd counting network.