| Summary: | 碩士 === 元智大學 === 工業工程與管理學系 === 98 === Object tracking is one of important applications for intelligent video surveillance. Traditionally, surveillance relied on human eyes. Error often occurs because of fatigue or psychological reason of person. Because of the advantage of automatic object tracking in accuracy and stability, object tracking is great solution for surveillance field to replace human eyes. The applications related object tracking prosper and related researches are launched worldwide.
This paper proposed a mixed object tracking method which combines two methods, SAD and RGB histogram, with dynamic weighting for different scenes to increase the tracking accuracy. The weightings are determined by backpropagation neural network (BPN). Four indexes of content of videos for representing the characteristic of the video and best weighting were selected as input and output of BPN. Large amount of video were utilized to train the BPN. Using the trained BPN, the method should get better performance than SAD or RGB histogram.
In the experiment of this paper, two kinds of videos, person and automobile, were taken to verify the performance of the mixed-ratio method. Mean square error (MSE) of position error for tracking within frames of video is the index of performance. The results show that the MSE of the mixed-ratio method proposed in this paper for all videos is the lowest among the three methods, SAD, RGB histogram, and mixed-ratio method.
|
|---|
