Low Quality Face Detection And Application On Vehicle Management Platform For "Two Kinds Of Buses And Vehicles Carrying Dangerous Goods"

In China,the "two kinds of buses and vehicles carrying dangerous goods" mainly include long-distance buses,travel chartered vehicles and special vehicles carrying flammable and combustible materials.In the surveillance for it,due to the factors such as uncooperative passengers,uneven illumination,noise and small face,the performance of the existing face detection algorithm is not satisfactory.This thesis studys the low-quality face detection algorithm in this situation.This thesis studys the low-quality face detection algorithm systematically for "two kinds of buses and vehicles carrying dangerous goods".This thesis first introduces the characteristics of surveillance image database,then preprocesses images and detects faces from coarse to fine,finally introduces the design and implementation of the vehicle management platform for "two kinds of buses and vehicles carrying dangerous goods".The main work of this thesis is as follows:1.Pretreatment of low quality surveillance images.In the aspect of de-noise,three methods are introduced,which are bilateral filter,median filter and non local mean denoising.The performance of the three denoising algorithms is compared according to the peak signal to noise ratio,the similarity measure based on structural similarity and the subjective quality evaluation criteria.As a result,due to the bilateral filter peforming well in denoising as well as preserving image edge characteristics,bilateral filtering is selected for denoising algorithm in this thesis.In image enhancement,this thesis compares three kinds of image enhancement algorithm based on Retinex.Due to the performance of MSRCR for color fidelity is better,MSRCR is a good choice.The quality of pretreated image is improved in a large scale.2.Rough face detection based on two order Gauss mixed skin color model.Experiments find that the complex background can cause bad interference to the face detection.Because the two order Gauss mixture model of H-SV and C'b C'r has a good perfermance of separating brightness and color,the skin color detection algorithm based on the model is more robust to illumination and has better skin color detection performance.This thesis uses this method for skin color detection,then the result of detection is morphological processed to eliminate holes in the face region,finally the face region and background region is divided.3.Face detection algorithm based on Haar-Like L features.For the problem of small faces,the Haar-Like L features is proposed to describe the local information of small face,because the face contains many kinds of L type structures,the L features describe the small size face well.Experimental results show that the proposed detection algorithm based on Haar-Like L features has better performance in small faces detection.The algorithm of deep learning face detection based on Tensorflow is studyed,experiments show that in the case of small positive samples,the algorithm proposed peforms better and is less time-consuming.4.Design and implementation for vehicle management platform.This thesis introduces the system architecture and development tools firstly,then introduces the function modules and the design of database,finally,introduces the optimization methods of database and the on-line process.Among them,the number of detected faces is the basis for judging whether the bus is overloaded.In addition,the dynamic data and static data are partitioned,and the statistical results of the day are saved for shortening the time to generate statistical reports.