Research On Human Detection Algorithm Based On Space Context

Human Detection is one of the rising research area on image process focus theseyears. Nowadays, the comparative maturity focus of object detection is face detection,pedestrian detection and so on. The former detection of human body is that running onthe still image on which human is on the fixed posture, meanwhile, the detection rateis a little lower. Human detection has a high valuable application, such as it can beapplied to controlling traffic to reduce the traffic accident, and it can help theintelligent car driving, it can also be applied to frontier defence.These years, more and more nature disaster has taken placed on some countriesand areas. So many natural disaster have taken placed on the peacefull earthconstantly, that people died for these disaster are innumerable every year. Therescuing work has been the significant job concerned by the government. Many ofthe disaster scene are unreachable and dangerous. Currently, many scientists havedevoted themselves to reasearch and development the rescuing equipment and robothoping to help the aid worker.The paper proposed a kind of method based on space context of image. It is tohelp the project of Country863program “The research of perception technology onthe robotic for searching victims on the disaster scene”. Space Context is a conceptof human behavioural science, which is similar to the concept of attention mechanism.The space context mechanism proposed in the paper is maily showed in the validationof detected windows. The first step of the detection took HOG (Histogram of Gradient)operator as human feature, applying AdaBoost classifier which has goodperformance on classifying to machine learning, then validated the detected windowsbased on space context. Considering the problem proposed above, the paper presentthe improved scheme to reduce the detection speed and false alarm. ConventionalHOG operator is based on dividing image into block, on which the gradient iscaculated on the unit of cell. So the quantity of feature vector on each image windowis much large,which need long time to classifiying learning, and the detection timecan’t reach to real-time when the ratio of detection area is large. The paper improvedthe HOG operator that the image is directly divided into cell to calculate the gradient.It took relevant experiment to get proper rate of coverage when extracting the featureso that to decrease the detection time while keeping the detection rate same as before.Besides, it added the validation of human based on LBP texture to the procedureof space context. It applied LBP(Local Binary Pattern) operator to get human texturefor the LBP classifier, in viewing that the diversity of pose, the paper selected theuniform pattern LBP operator, took SVM(Support Vector Machine) which has goodperformance on classifying small amount of samples. Then it applied the producedhuman texture classifier to validate detected window. The result shows that this algorithm can exclude the false alarm produced by complex background effectively.The experiment of the paper taking shows the human detection algorithm basedon space context can detect the victims on various posture, such asstanding,sitting,laying,and standing with occlusion,laying with occlusion and so on bysingle human model. Meanwhile, the detection rate is higher than other methods andthe missing rate is low. The validation based on space context can reduce the falsealarm rate effectively.