Research On Pedestrian Recognition Algorithm Based On Video Sequence

Nowadays,pedestrian recognition technology has important commercial value and application prospect in the fields of intelligent monitoring system,driver assistance system and intelligent transportation.However,pedestrian posture,movement,scene occlusion and changes in lighting and other factors will have an impact on the accuracy of pedestrian recognition.This thesis aim these problems to improve the accuracy of pedestrian recognition.The main research contents are summarized as follows:(1)Based on Support Vector Machine(SVM)for pedestrian recognition: linear SVM is used as classifier,Gabor wavelet and Histogram of Oriented Gradient(HOG)are fused as sample features.Firstly,apply Gabor wavelet transform the original image,then the result image is fused in direction and scale to obtain a Gabor feature image.Based on this,the HOG features are extracted from the feature images to complete the G-HOG feature extraction.Finally,the SVM is used to recognize the pedestrian.(2)Pedestrian recognition based on two-dimensional BP neural network: Many data are usual y expressed in the form of matrix.However,traditional feedforward neural networks(FNNs)are based on vector input and must be decomposed into vector forms when used.As a result,the information between the original matrix easily lost.Therefore,we proposed a two-dimensio na l backpropagation algorithm(2D-BP)to train two-dimensional feedforward neural networks(2DFNNs)model for pedestrian recognition,and adopted stochastic gradient descent method to learn all the weights in this network.Finally,the model learned by this algorithm effectively preserved the original structure of two-dimensional matrix data,it is favorable for pedestrian recognition.(3)The experimental results and analysis of two kinds of pedestrian recognition algorit hms : The comparison between pedestrian recognition algorithm based on SVM and other algorithms in pedestrian database shows that G-HOG features of extracted images can better characterize pedestrians and the obtained classifier by SVM shows a higher recall rate in the database of pedestrian posture and background diversity.The pedestrian recognition algorithm based on twodimensional BP neural network demonstrated the experimental parameters on a standard database,and compared with other algorithms.The result shows that the algorithm can select the strong features of discrimination ability from a large number of training samples,After adjusting the parameters,the model is more stable and shows a higher accuracy on the database.(4)Pedestrian recognition in video sequences: By analyzing the experimental results of two pedestrian recognitio n algorithms,it is found that for pedestrians in the video sequence,a pedestrian recognition algorithm based on a cascade classifier can be used.Firstly,the SVM classifier is used to detect and store the pedestrian suspicious area,and then the two-dimensio na l BP neural network is used to further identify and improve the accuracy.The experimental results show that the proposed algorithm can effectively improve the accuracy of the recognition in both the pedestrian database and the video sequence captured in the actual scene.