基于场景信息注意模型的目标检测技术研究

Research on Target Detection Technology based on Scene Information Attention Model

笔者为克服自底向上的注意模型的不足,加入了自顶向下的上下文静态和动态场景信息。场景信息能更好保留有用信息、去除噪声干扰。笔者运用中央-外围差分算法,并且运用高斯分解进行多特征融合生成总显著图。在特征图中进行显著目标的提取、跟踪及焦点转移。通过引入上下文场景信息的概念,结合图像的动态和静态特征,使注意顺序更符合人类视频注意机制。实验结果表明,测试视频中的跟踪焦点总是集中在显著性最强的显著区域中。

In this paper,the author adds the contest-aware saliency that contain the static and dynamic information based on top-down visual model to improve the performance of bottom-up visual attention model.The algorithm of central–surround can combine multi-features to generate the finally saliency map in the Gaussian pyramid which could reduce the noise impact.In the feature map,the significant target will be selected and then transferred.By using the contest-aware saliency map,the tracking focus is always concentrated on the visual saliency area of the moving target.The experimental results show that the visual attention model is fit to the mechanism of the human attention.