The Detection And Tracking Of Human Based On Video-sequences

In the perception of outside information, visual information account for a large of proportion. In particular, dynamic visual information is the main component. In the environment, the perception of these dynamic information s extraction, detection and identification has become an important computer vision research, which has been concerned by the domain of Computer. At the same time, with the computer vision and image processing technological development, the processing capacity of videos has been improved greatly. It makes the subject of the detection and tracking of human based on video-sequences has been developed greatly. From the development of the Monitoring system, it has experienced three periods: simulative scene monitoring system, digital scene monitoring system, intelligent scene monitoring system. From the functional view, surveillance system Video-based can be used in auto video s surveillance, bank and mall s surveillance, large-scale public facilities Monitoring, Medical care and Human-Computer Interaction(HCL) Monitoring system etc. From the applications view, the range of the application has developed from military institutions, customs, financial Institutions, airport monitoring to residential, office, Parking Monitoring etc. which can be seen anywhere. Therefore, the topic of this research has an important theoretical significance and practical value.The main task of the detection and tracking of human based on videos-sequences is to detect the existence of moving regions from the background; and then to detect and identify the human from it. Finally, it tracks the movement of human in the video-sequences. The combines of human identification and tracking algorithm make tracking the movement of human more effective, which save a lot of time. There are many difficulties in this subject both in theory and in practical applications we have faced with. For example, how to ensure the accuracy of the region which has been extracted, how to improve the accuracy of the algorithms of machine-learning; How to adapt itself to the changement of the size of the rectangles and the templates in tracking. Many experts and scholars around the world have carried out a lot of researches in this area. In this paper, I have studied mainly the algorithms of machine-learning in human detection and the algorithms of tracking on the basis of these studies.On the one hand, this paper research and analysis the use of Adaboost training algorithm in human detection may have two problems: the over-adaptation of target weights and the degradation of training, and make a series of improvements. In the simulation experiment, which has verified the algorithm s effectiveness and precision. On the other hand, this paper has completed the integrate of the Kalman filter-based motion estimation and human tracking region-based algorithm, which has solved the tommyrot of separate human area tracking algorithm. It can not get good tracking results in the presence of the moving target s block problems. In the simulation experiment, it shows that this combined method in dealing with the over-adaptation of human occlusion has achieved good results. Thereby it reduces the probability of Missed tracking . The principal works are as follows:1,Proposing the concept which is about the error ratio of positive and negative samples in the paper. We can control the distribution of the target weight which may lead to the over fitting phenomenon of the traditional Adaboost algorithm by controlling the error ratio of positive and negative samples. In the Adaboost training process, the positive training samples was classified as negative samples in the classification shows that the corresponding features of the classifier are hidden or not obvious. That belongs to missing detection phenomenon and be recorded as positive error ? ?; the negative training samples was classified as positive samples in the classification shows that the corresponding features of the classifier are positive features. That belongs to false detection phenomenon and be recorded as negative error ? ?. Then we can calculate the value of? . The improved Algorithm reduces the possibility of over fitting phenomenon about the distribution of target weights, which will raise the efficiency of the algorithm.2,Increasing the sensitivity to the wrong times of samples in the process of updating the weight, that makes the weights of tough samples which have been misclassified several times doesn t to become too large. It is treat the whole wrong samples as "the same one" by the traditional Adaboost algorithm. One of drawbacks is that it is not benefit to control the weight of the samples which have been misclassified several times. At the same time, the weight of the samples, to some extent, which have been misclassified fewer times also be controlled. In order to overcome the drawbacks above, it was shown that the change of k of those samples above has the opposite trend to compare with the weights expansion. The higher k is, the lower the weights expansion. The improved weight update algorithm can limit the growth of the weight, and the weight s distribution is more reasonable than before. The degradational phenomenon doesn t occur.3,Changing the forms of the final strong classifier s. In the Adaboost training process, recording the priori probability of each weak classifier, removing the implicit features. In the detection process, we can use the priori probability to guide our classification to these images to be judged. That will output two probabilities: the probability of the positive samples and the probability of the negative samples. However, the sum of them is not equal to 1. With the improvement, it makes the strong classifier more reasonable to judge that images, increasing the algorithm s robustness.4,Completing the integrate of motion estimation algorithm based on Kalman Filter and Regional Track in the human tracking process. When two movement regions don t to be connected, we can use the Regional tracking algorithm. On the contrary, we can use motion estimation algorithm based on Kalman Filter by calculating the gravity center. In tracking, it only changes the location of the target area, and does not change the size of external rectangular box. The algorithm can be well used to solve the situation of mutual goals Occluded problems.