Video Pedestrian Detection Based On Deep Learning

Pedestrian detection technology is one of the key technologies in the field of computer vision and image processing.It has immeasurable application value and prospect in urban intelligent transportation,virtual reality and human-computer interaction.Pedestrian detection technology based on video images and static images are the two research focuses of pedestrian detection.This paper mainly studies the detection technology based on video images.The paper firstly introduces the background and significance of the research on pedestrian detection technology,analyzes and summarizes the technologies and algorithms involved in each part from the extraction of Regions of Interest(ROIs)and pedestrian detection algorithms.Then the paper introduces the principle of traditional pedestrian detection algorithm and the principle of pedestrian detection algorithm based on deep learning.In the traditional pedestrian detection,three steps are summed up from the extraction of the image ROIs to the training of the ROIs features and the feature classification.In deep learning pedestrian detection,the paper concludes the bottom-up detection process from data set to model selection,model testing,model optimization and model detection.Next,based on the basic theory of pedestrian detection algorithm,this paper presents two improved algorithms.(1)The paper studies Co-occurrence Histograms of Oriented Gradients(CoHOG)features and improves CoHOG features from the perspective of feature extraction.The paper first introduces the CoHOG principle and points out that CoHOG failed to consider the gradient amplitude information and the slow detection speed caused by large amount of feature calculation.In order to solve this problem,the paper proposes a CoHOG feature of weighted amplitude(WA-CoHOG)and chain cascade classifier based on weak classifier sorting algorithm.The gradient information is introduced into the WA-CoHOG features by weighting function,and the small features of each block are obtained by segmenting the image.Finally,a small number of features are input into the cascade classifier to calculate the classification results.Simulation results show that WA-CoHOG features can make full use of the gradient magnitude and direction information to improve the pedestrian detection accuracy.The cascade classifier can shorten the detection time effectively.The algorithm proposed in this paper can get faster detection speed while getting higher classification accuracy.(2)This paper studies video pedestrian detection algorithm based on Graphics Timing.In view of the shortcomings of the pedestrian image timing and context information that can not be extracted and utilized in the video pedestrian detection,A video pedestrian detection network is proposed for improved deep graphics(IGT)deep learning.The network uses multiple contextinhibition methods to sort all the recommended box detection scores in descending order.The high confidence box with a score above the threshold will be retained,while the low confidence box with a score below the threshold will be subtracted from a small value to suppress it.At the same time,using the information vector to propagate the optical flow of all the pixels,the average optical flow vector of the bounding box is calculated,and the frame coordinates with the same detection score are propagated to the adjacent frames according to the average optical flow vector to suppress the false negative detection.Finally,the network effectively improves the detection accuracy through end-to-end training.