Research And Implementation Of Pedestrian Detection Algorithms Based On Machine Vision

Pedestrian detection is an important technology in computer vision and pattern recognition.It has high application value in intelligent transportation,driverless,intelligent tourism and other fields.Its research purpose is to locate pedestrians accurately in image or video sequence,but real-time pedestrian detection has certain challenges,because pedestrians are easily affected by illumination,occlusion,background and other factors.It is of great significance to design an ideal and real-time pedestrian detection method.The core issues of pedestrian detection mainly include two aspects: improving the detection rate and speed.From these two perspectives,the contents of this paper are as follows:(1)An improved MLBP feature and CMLBP feature based on uniform pattern LBP feature are proposed.Firstly,an improved texture feature MLBP(Mean of Local Binary Pattern)and CMLBP(Color Based on Mean of Local Binary Pattern)based on uniform pattern LBP feature are proposed.The improved MLBP and CMLBP features are fused with HOG features and CSS features,and the optimal feature is selected to improve the detection rate of pedestrian detection.The experimental results show that the detection rate of MLBP feature is 3.5% and 2.1% higher than that of uniform pattern LBP feature and basic pattern LBP feature,respectively.The detection rate of combining CMLBP feature with HOG feature is 95.25%.(2)An improved feature HWEBING based on BING features is proposed.Based on the idea of BING features to improve detection speed,this paper proposes an improved feature HWEBING(Hash and Window Enhancement of Binarized Normed Gradients)based on BING features to pre-detect the image and improve the speed of pedestrian detection.Pre-detection can reduce a large number of non-object windows and screen out candidate windows which may be objects,thus greatly improving the detection speed.After using HWEBING features to pre-detect the candidate window,MLBP and HOG features are extracted from the candidate window to improve the detection rate of pedestrian detection.The experimental results show that the proposed method is 5.5 times faster than the traditional multi-scale pyramid scanning method,and after using HWEBING pre-detection,the detection rate of extracting HOG features from candidate windows is higher than that of using BING features for pre-detection.(3)A pedestrian detection method based on image enhancement is proposed.Finally,a pedestrian detection method based on image enhancement is proposed to improve the detection rate of pedestrian detection.Before extracting features,the image of the region of interest is interpolated to enhance the pedestrian texture information in the image.LBP feature of gray image loses information on color channel.After image enhancement,IE-CLBP(Image Enhancement-Colorized Local Binary Pattern)feature based on color space is extracted,which describes the texture information of image more comprehensively.The experimental results show that the detection rate of LBP,CLBP,HOG + LBP,LBP + CSS extracted from the enhanced image is significantly higher than that of the original image combined with SVM classifier and HIKSVM classifier.In practical application scenarios,pedestrian detection is often affected by background and occlusion.It is difficult to meet the real-time and stability of pedestrian detection by using pedestrian detection alone.The combination of pedestrian detection and pedestrian tracking method can be better applied.Therefore,an optimized pedestrian tracking algorithm is proposed to improve the tracking accuracy:KCF(Kernel Correlated Filter)algorithm can easily lead to tracking failure when the target is occluded in the tracking process.So the HOG feature and the improved LBP feature(MLBP feature)are extracted in the training stage.Because the two features are complementary,the fusion of the two features has better robustness to occlusion.The experimental results show that the improved KCF algorithm has higher tracking accuracy than the original KCF algorithm.