Research On Pedestrian Detection Based On Deep Learning Theory

In the current target detection field,pedestrian detection has an irreplaceable status.It is both a difficult issue and a hot topic for researchers.In recent years,due to the rapid development of artificial intelligence,the mature pedestrian detection technology has been further requirement in accuracy and speed.Deep learning is a hot issue in the field of machine learning,so the application of that on pedestrian detection is a particularly important research direction.This paper focuses on the method of pedestrian detection based on machine learning.Resorting to the basic theory,it detailed describes the pedestrian's feature representation and classification strategy and systematically summarizes the key difficulties in current pedestrian detection problems and the problems remained in existing methods.The simulated results show that the traditional HOG+SVM and CNN-based pedestrian detection methods have some defects,such as complex pedestrian characteristics and large amount of calculation.Furthermore,the detection results are easily affected by such factors as occlusion,complex background,and illumination and other factors.Therefore,this paper proposes an improved CNN model,which mainly involves in two stages as follows.(1)Firstly,in view of the performance of existing CNN-based pedestrian detection algorithms,it is found that they all directly perform convolution operations on the original images which makes large amount of calculation and the detection results are easily effected due to complex backgrounds.Hence,we propose to firstly treat the picture for preprocessing,filter out some useless features,and then perform convolution operations on them.In this paper,gradient pre-processing and texture pre-processing are respectively adopted to on the inspection images.The detection performance of the algorithm under different pre-processing operations is analyzed.Experiments show that the detection performance after texture pretreatment is better than that after gradient pre-processing.(2)On the basis of the previous text,the selective search algorithm is used to extract pedestrian preselected frames from the images to be detected,and the sliding window method brings about a large number of region proposal;the extracted region proposal are then scaled up and unified.A grayscale image with a size of 64*128 is used to perform texture preprocessing on the scaled image.The complex background is filtered out to highlight the texture features of the pedestrian.Finally,the texture-preprocessed feature map is input into the CNN model for detection.The experimental results show that the improved CNN model in each picture False Positive Per Image(FPPI)takes accounted for 0.1%of the total positive cases,and the missed rate is 27.2%,compared to that of the classic CNN 38.4%with a decrease of 11.2%.