Research On Algorithms With Applications For Real-time Hand Gesture Detection And Tracking

With the rapid development of modern science and technology,especially the computer science,communication,electronics and sensor technology,control theory and artificial intelligence,human-computer interaction has been widely applied to the scientific research,industry and agriculture,military affairs,medical and health care,transportation and so on.Gesture-based interaction employs gestures to express interactive intention,which contains plenty of interactive information and accord with the mental model of human.Therefore,research on human-computer interaction(HCI)based on gesture has become the current focus of attention.However,since the hand is not rigid and is distorted easily,the failures to track often occur if these tracking alogorithms are used,affecting man-machine interaction.In addition,the present algorithms lie in the shortage in removing the background,especially complicated background.Although Kinect Box 360 launched by Microsoft has a good performance in HCI,it is too expensive to be used by intelligent home appliances.Therefore,research on HCI based on universal cameras not only has the important theory significance,but also has the important economic value and applied value.In this dissertation the selective topics are focused on algorithms and corresponding implementing technologies of image denoising,gesture detection and gesture tracking.The main achievements and innovative contributions are following as:(1)A novel adaptive denoising method based-on PCNN is proposed to implement the removal of salt and pepper noise effectively and the conservation of image details and textures,in which a kind of detection mechanism is applied to discriminate whether a given pixel was corrupted or not,and only those corrupted pixels must be denoised so that the original image information can not be damaged and the details and textures of images can be conserved effectively.In order to improve the image quality,the self-organization mechanism is introduced into PCNN array framework,thus neighboring connection modes of neurons in the PCNN can switch automatically.Furthermore,an adaptive window mechanism is also used to automatically select the optimal filtering times based on the estimated noise intensity so as to enhance the adaptability of algorithm.Experiment results indicate that this method presented is more preponderant than the conventional methods and other congeneric methods in removing noise and conserving image details.(2)In order to achieve gesture detection with background independence,a gesture classifier using Cascade Ada Boost is designed.It is well-known that it is a difficult problem for image processing to extract gesture from complex background.The traditional methods,for example,background difference,background difference based on Gaussian mixture model,frame difference,can not meet the practical application needs.The background difference method or Gaussian mixture model-based background difference requires background is not changed,but actually the background is often changes.Therefore,this method is usually only used in the laboratory environment or very strict requirements of environment.The frame difference also has two problems: the first,it requires only a motive object in the video,the second,it require that illumination maintain a constant.But for the intelligent television users these requirements are difficult to be ensured,in addition,the method can not extract the gesture object precisely,which will affect the extaction of tracking feature points,thus leding to the failure to tracking.It is very important to put forward a method with background independence for gesture extraction In view of this,with the aid of face detection,a classifer based on Cascade Ada Boost is designed through a large number of training samples.The experimental result show that our method is highly efficient,accurate and real-time.Moreover,it has very strong robustness and low false alarm rate.(3)This paper proposes an adaptive target tracking strategy and solves the disadvantages of classical KLT algorithm.The experimental results show that the algorithm has a high tracking accuracy,the reliability and robustness are significantly better than the KLT algorithm and Cam Shift algorithm.This is mainly attributed to the following strategies: filtering strategies: reducing the effect of the noise and light changes on KLT tracking algorithm;forward and backward error strategy and reliability metrics: reducing the wrong tracking;the mechanism of adding feature point tracking: decreasing the impact of the number of feature point reducing for the deformation.