Research On Pedestrian Detection And Image Clustering Based On Feature Fusion

With the rapid development of artificial intelligence and computer performance,image data processing has become the focus of research in the field of computer vision.Pedestrian detection technology can be used in areas such as traffic monitoring,human-computer interaction,intelligent driving,behavior analysis,etc.It has an extensive applied foreground.However,in real life,pedestrian detection are often affected by many factors such as illumination changes,deformation,occlusion,etc.Limited pedestrian data sets and the low efficiency of manual annotation result in difficulty in extracting features manually.Pedestrian detection becomes more challenging.In addition,human accumulate a large number of image data,but these data are unlabeled,the efficiency of manual annotation is low,and the effect of traditional clustering algorithm is poor.To address these issues,the main research work of this paper includes the following:(1)A pedestrian detection method with multi-feature fusion is implemented.The single feature of image can not accurately describe the pedestrian,so multi-feature fusion is used to extract pedestrian feature.This paper extracts their HOG(Histogram of Oriented Gradient,HOG)features and Uniform LBP(Uniform Local Binary)characteristics in key areas,such as head and limbs,then sets the weight coefficient according to the strength of the identification ability of two features,and the feature vector is combined by adaptive weighting fusion.Finally,the SVM(Support Vector Machine)classifier is used for detection.On the public pedestrian image dataset,the algorithm is used to experiment.Experiments show that the proposed method based on image segmentation is more accurate and faster than the traditional method which only uses single feature.(2)A sparse convolutional neural network pedestrian detection algorithm based on multi-channel is proposed.The features of artificial extraction generally start from the original data,and it,which need a thorough analysis of the data,requires the help of experts' domain knowledge.It also costs much manpower and computing resources.Based on the structure of the convolutional neural network,a better pedestrian detection algorithm with better effect is constructed in this paper.It is unnecessary to manually design features and the method can automatically learn the characteristics of the data.The input of the traditional convolutional neural network is the original image,so there is much redundant information and the local edge and texture are not clearly characterized.This paper proposes a sparse convolutional neural network detection algorithm based on multi-channel.Firstly,the HOG feature map and YUV color space are formed into three channels.New feature is extracted through the convolution layer,and the sparse automatic encoder is used to sparse the feature from neural network.Finally,the softmax classifier is applied in pedestrian detection.The model makes full use of the pixel-level features of the image and the advantages of the HOG feature in describing the contour representation of the pedestrian,and improves the detection efficiency of the algorithm.(3)A deep convolution clustering algorithm based on feature fusion is proposed.Due to the lack of labeled samples in real life and the high cost of manual labeling,the feasible method is to let the machine automatically label images,which can prepare data for image classification tasks such as pedestrian detection.Traditional deep clustering algorithms do not fully exploit the advantages of convolutional neural networks and cannot protect the local structure of the original data.Based on DEC(deep embedding cluster),this paper uses convolutional autoencoders to learn features and defines a new loss function for training to ensure that the clustering center of each class is as far away as possible.Simultaneously the output of the coding layer is fused with the output of the previous layer of it to form a clustering feature that needs to be optimized by cluster loss.This can reduce the loss of feature information and improve the effect of cluster.