Research On HOG Fusion Feature And DL In Pedestrian Detection Algorithm

In the computer field,pattern recognition has always been a challenging subject.Although there are many obstacles and bottlenecks,it still attracts a large number of experts and scholars.Pedestrian detection is an important area of pattern recognition.It has been widely used in intelligent monitoring,human-computer interaction and danger warning.This paper studies the existence of pedestrians in static images.The paper firstly introduces the research background and significance of pedestrian detection,then analysis the current research situation at home and abroad,and briefly summarizes the feature detector,classifier and commonly used models of deep learning.According to some shortcomings existing in research technology,this paper carries out the work from two aspects:(1)Recently,most detection technologys focus on how to improve the characteristics and classification.So,an improved fusion feature algorithm is proposed for pedestrian detection in this paper.The algorithm firstly performs gray-scale processing on the input image,and then extracts the HOG feature and performs second-order wavelet transform.It reserves the low frequency sub images(LL,LL1)obtained by the two stage decomposition,and calculates the LBP feature spectrum by using LBP detector,then calculates the histogram.The wavelet LBP feature is obtained.Merging the HOG feature and the wavelet LBP feature and reducing the dimension by using the Principal Component Analysis(PCA),the final fusion feature is obtained.Put the final fusion feature into the Support Vector Machine(SVM)to train and test.And the experimental results show that the proposed method has certain advantages in the contrast algorithm.(2)Through analysing and understanding of the deep learning,and combining the advantage of the autonomous learning feature,the paper uses the convolution neural network to extract the image features,predicts the preselected area box in the image,and makes the preselected area network and detection network realize the feature sharing.This can reduce the computational complexity greatly and reduce the detection time.On this basis,the paper carries out a large number of experiments: first,verify the validity of the pre-selected area network,then fine-tunes the threshold.It is clear that the locating accuracy is improved obviously.At the end of this paper,we verify the feasibility of the algorithm on ImageNet dataset,INRIA dataset and self-made dataset respectively.