Human Detection System Based On ADABOOST Algorithm

Human detection system,which involves the advanced technologies in the fields of computer vision,image processing,pattern recognition,computer science and so on,has developed into the significant field of computer visual research.The thesis focuses on the technologies of moving human detection in a static situation and then give the implementation of the system.The implementation of human detection system mainly includes two parts of moving object detection and human detection.In the part of moving object detection,we briefly introduce some of the common algorithms,including background subtraction method,frame subtraction method,background modeling method and introduce the frame subtraction method in a detail. At the same time,we introduce the algorithm of morphology treatment and extraction of connected which reach the goal of optimized treatment to the contours of moving object.To the part of human detection,we mainly focused on object recognition based on the Adaboost algorithm,including the calculation of haar features,training of weak classifier,iterative selection of weak classifier,training of strong classifier and the construction of cascade classifier using strong classifiers at all levels.The thesis gives an overall summary of the rationale of technology of video image segmentation and presents a implementation flow of moving object detection and segmentation and proposes a modified calculation based on the traditional,the frame subtraction between three frames.The virtue of traditional method is that it sensitive to moving objects only and in fact it can only detect objects in relative movement.It has a valid and steady results on account of tiny impact of light result from short time interval between two images.It is a pity that this method often brings undesirable consequences:one is that it can not detect out the overlap between two frames which result in a cavity in the object;the other is that the position is not accurate.The external rectangle at the moving direction is stretched which results that the detected object is bigger than the real aim.This is in fact because that the relative motion and object location is not exactly the same.Relative motion is effected by the object's own velocity and the time interval between successive image.The thesis proposes a modified method which based on subtraction between three frame images can preferable detect moving target from sequence of images.Although it still has hole in the object,it solves imprecise problem of detection of moving object which exits in traditional method.It can better suppress noise and eliminate interferer brought by the change of background.The weakness of this method is that it has a frame lag on treatment.But the system basically reach the requirements of real-time because of the sampling and transmission speed by current video capture device is more than 15 frames per second.The thesis also makes some improvements on consideration of request of real-time and the raise of detection accuracy of the human detection system.It mainly reflect in two sides:one is to filtrate detection region which is ready for detecting.It is to compare all the moving region to the testing window size,and than eliminate the regions which have smaller size than the resting window size.This step is complete before coming into the Adaboost classifier to reduce the computation and to speed up the detection rate.The other is to predict the enlarge scale of testing window size. After computing detection regions' biggest size,we can obtain a ratio by calculate the specific value of height and width between the biggest size and the testing window size.And the ratio is the right time to amplify.It can avoid undetected and repetitious calculation.In the aspect of human detecting,the thesis gives the implementation process of human detecting based on Adaboost algorithm.It gives several common haar feature template,then compute haar feature using the integral image.Thousands of weak classifiers whose performance are as low as about 50%are formed by computing the haar feature of each detected moving targets.We train the better weak classifiers that are selected by iterative algorithm to form all levels of strong classifiers.So we can say that each level of strong classifier are formed by different number of weak ones with the purpose of making the strong ones' accuracy higher step by step.At last,we combine all levels of strong classifier to a cascade classifier through some algorithms whose classified accuracy has reached more than 90%and with better robustness.Recently,moving human detecting problem has become a hot topic in the field of computer vision.It was also seen as an important capacity of computer vision systems and has a wide range of applications in the fields of artificial intelligence,military affairs,industry and so on.In these aspects' research work have achieved amazing achievements.For the aspects of moving object detection,optical flow method,frame subtraction method and background modeling method have been addressed which can not only solve object detection problem in static condition,but also the optical flow can solve the dynamic change situation which has a more complex implementation and larger calculation combined with bad real time and practicability character.In pattern recognition,the systems of handling single person,multi-person and overlapping situation have been accomplished.The correlation algorithms based on data clustering,statistical classification,neural network and structural pattern recognition have been brought up.Especially the algorithms based on SVM and Adaboost in the field of statistical classification have formed natural theoretical foundation which have been widely used in the actual systems and get a perfect performance.The experimental results show that the human detecting system with the three frames subtraction,human detecting based on Adaboost algorithm and the modification for the demands of real-time system and high accuracy has a good performance.In the static situation,the accuracy of moving human detecting is very high and it also has a good robustness.The system has basically meet the requirement of moving human detection.