Detection And Tracking Of Moving Object In Video Series

In the last decade, due to the progress of multi-media technique and computer performance, Moving Target Detection and Tracking (MTDT) based on Dynamic Image Sequences has been developed rapidly, and has become an extremely popular field in computer vision. Now, MTDT is applied widely, such as Intelligent Monitor, Intelligent Robot, HCI, Virtual Reality, Medical Diagnosis, Video Compression and so on. Actually, most meaningful vision information is taken in the state of moving. Therefore, moving target is naturally becoming research focus so as to endow the study of MTDT with an extraordinary realistic and applicative value. Based on the result of the classification and comparison of the existing methods of MTDT, improved algorithm for MTDT is presented in this thesis.Four common MTDT algorithms are discussed, and a method of moving target detection which includes the adaptive background model and the dynamic threshold background distinction algorithm is proposed in part 1. Firstly, the algorithm introduces the Surendra method to update vision background, and then accomplish the real time detection by the final threshold, which is calculated by adding the threshold of Ostu algorithm to an increment reflecting light fluctuating. The present algorithm uses a new background observation data to update adaptively the background model in order to improve the flexibility and stability of the system, and uses the methods of the shadow detection and noise removing to increase the veracity of detection for a moving target. Multi-group image sequences, which are taken in natural environment, are used to validate the algorithm of body-moving detection. The results indicate that the new algorithm has a good efficiency and accuracy.Target tracking methods are summarized in part 2. Firstly, the applicable scope of three existing forecasting algorithms are analyzed and the question on general target tracking in the frame of Bayes is discussed. And then, the algorithm for Monte Carlo integral and the particle filtering algorithm based on sequence significance sampling is demonstrated in detail. The algorithm forms a particle filter model of target tracking to estimate the moving state of target. Therefore, the noise interference and the scope for feature extraction has been decreased, which helps to improve the efficiency of the present algorithm. In addition, the modified method of particle filtering is shown in this thesis, which can real-time track the moving target in the complex environment. The modified method initializes the particle filter again to solve the problem of inaccurate tracking caused by invalidation of moving model in the whole process, and proofreads both the detection and tracking results to improve its stability. Finally, the experiment results manifest that particle filtering algorithm can solve the problem of non-linearity and non-Gauss estimate, as well as the tracking algorithm based on particle filter can track target in the complex environment.In the end, the conclusions of MTDT are reached, the existing deficiencies and the direction for further research are discussed.