Research On The Online Pedestrian Detection For Intelligent Vehicle In Foggy Environment

Pedestrian detection is one of the key technologies of intelligent vehicle's environmental perception system.Because of the diversity of pedestrian posture and the complexity of background,it's difficult to detect.In this regard,this thesis proposes an algorithm of F-DPM based on the DPM,which could realize fast and accurate pedestrian detection.In addition,foggy weather will degrade the image quality of camera,which will affect the result of pedestrian detection afterwards.Therefore,a defogging preprocess is carried out by the Refined Dark Channel defogging algorithm,which improves the image quality of fog image.The specific working arrangements are as follows:Firstly,the theory and implementation process of HOG feature extraction are studied,and the feature structure is defined.Based on this,a DPM pedestrian detection model is constructed.Through the soft connection of root filter and component filter,the effective detection of multi-pose pedestrian is realized.Through the construction of image pyramid,pedestrian at different scales can be detected effectively as well.As for the process of model training,the convergence speed and classification performance of SVM classifier are improved with the training method of hard sample learning.As for the process of detection,the F-DPM pedestrian detection algorithm is proposed based on the construction of fast feature pyramid and the Discrete Fourier Transform,in order to improve the detection speed.Secondly,defogging preprocess of fog image is realized,based on the Refined Dark Channel defogging algorithm.The image characteristic of fog image and fog-free image is studied,and the influence of foggy environment on image quality is clarified.Based on the incident light energy absorption model and scattering light scattering model,the degradation model of fog image is constructed.After that,the theory of Dark Channel defogging algorithm is studied,and the parameter of estimated transmittance t(4)(x)is filtered and refined by the Guided Filter.After that,the algorithm of Refined Dark Channel defogging is proposed,which could improve the expressiveness of edge and detail after defogging.Last of the part,the Edge Gradient Enhancement Ratio is used to verify the superiority of the Rfined Dark Channel defogging algorithm,as the evaluation criterion.According to Michelson contrast,the fog grade evaluation parameter _MC'is designed,which is used to determine whether the image is needed to be defogged.Finally,a pedestrian detection system is designed and verified by experiments,and a pedestrian data set is produced.Through off-line model training,a pedestrian model classifier is obtained,which has a better classification performance,the Precision is 94.98%,and the Recall is 83.61%.After that,it is verified that the proposed defogging algorithm could effectively improve the detection distance and detection effect,through the static environment experiment in foggy day.In the crowded scene of campus,the pedestrian detection system is tested by vehicle dynamic environment experiment,the average accuracy of the system is 92.16%,and the average detection speed is 98 ms/frame.In this thesis,a pedestrian detection system is constructed,based on the F-DPM and Refined Dark Channel algorithm.This system could realize pedestrian detection effectively,and the complex environmental factor of fog is taken into account.There's an important practical reference value for improving the environmental perception system of the intelligent vehicle.