Abstract: Pedestrian detection is utilized as the fundamental of various computer vision applications.A typical and effective solution of pedestrian detection is combining histogram of oriented gradient (HOG) with support vector machine (SVM).In this paper,we proposed a novel pedestrian detection method,which uses convolutional neural network (CNN) in HOG＋SVM to obtain more comprehensively feature description through a variety of convolutional kernels in CNN.Firstly,it extracts shallow features from data set by CNN,and uses CNN’s characteristics of displacement,and scale and deformation invariance.Then,it merges strongly relevant features by analyzing correlation coefficient of shallow features.After this,it extracts HOG features from CNN shallow features which are weakly correlated with each other.Finally,it uses SVM to complete the training and classification.The experimental results show that the proposed method can obtain higher accuracy than the existing method in pedestrian detection.

Key words: Pedestrian detection,Convolutional neural network,Multi-convolutional features,Histogram of oriented gradient