针对视频序列维数高、帧间相关性大、运动轨迹复杂的特点,将LLE非线性降维算法用于视频处理,并重点研究了如何利用该算法对目标跟踪过程中的模板进行预测更新.由于单步预测方法在运动目标发生部分或全部遮挡时无法保证跟踪的准确性,进一步将时间序列模型与BP网络相结合实现跟踪目标的多步预测,从而可以弥补时间序列模型在单步预测方面的不足.实验证明,该算法能保证在运动目标跟踪过程中的准确性和鲁棒性.%Aiming at the features of video sequences, i.e., the higher dimension, larger relativity of frame, and complex trajectories, this paper proposed applying the reduction algorithm of LLE nonlinear dimensionality to video processing.In particularly,this paper focused on how to utilize the above algorithm to predictively update the model of moving objective tracking.Because the single-step prediction could not guarantee the accuracy in the complex environment with part or the whole hided,this paper integrated the time series model with BP neural network to achieve multi-step prediction, which could overcome the shortcoming of time series model.The experiment results show that this proposed method can attain better accuracy and robustness for moving object tracking.
展开▼
