Abstract: Aiming at the problems of simple fusion of spatio-temporal information and ignoring moving information in moving object detection based on visual saliency, a moving object detection method based on the dynamic fusion of visual saliency and moving information was proposed. Firstly, local and global saliencies of each pixel was computed by spatial characters extracted from an image, then the spatial salient map was calculated combining those saliencies by Bayesian criteria. Secondly, with the help of structured random forest, the motion boundaries were predicted to primarily orientate the moving objects, by which the motion boundary map was built. Then, according to the change of the spatial salient and motion boundary maps, the optimal fusion weights were determined dynamically. Finally, moving objects were calculated and marked by the dynamic fusion weights. The proposed approach not only inherits the advantages of saliency algorithm and moving boundary algorithm, but also overcomes their disadvantages. In the comparison experiments with the traditional background subtraction method and three-frame difference method, the detection rate and the false alarm rate of the proposed approach are improved at most more than 40%. Experimental results show that the proposed method can detect moving objects accurately and completely, and the adaptation to scene is promoted.

Key words: moving object detection, visual saliency, structured random forest, motion boundary, dynamic fusion

摘要： 针对基于视觉显著性的运动目标检测算法存在时空信息简单融合及忽略运动信息的问题，提出一种动态融合视觉显著性信息和运动信息的运动目标检测方法。该方法首先计算每个像素的局部显著度和全局显著度，并通过贝叶斯准则生成空间显著图；然后，利用结构随机森林算法预测运动边界，生成运动边界图；其次，根据空间显著图和运动边界图属性的变化，动态确定最佳融合权值；最后，根据动态融合权值计算并标记运动目标。该方法既发挥了显著性算法和运动边界算法的优势，又克服了各自的不足，与传统背景差分法和三帧差分法相比，检出率和误检率的最大优化幅度超过40%。实验结果表明，该方法能够准确、完整地检测出运动目标，提升了对场景的适应性。

关键词: 运动目标检测, 视觉显著性, 结构随机森林, 运动边界, 动态融合