面向目标检测的多尺度运动注意力融合算法研究

Research on Multi-scale Motion Attention Fusion Algorithm for Video Target Detection

运动目标检测是视频分析领域的关键技术之一,针对目前全局运动场景下目标检测算法的局限性,该文提出一种多尺度运动注意力融合的目标检测算法,为目标检测问题提供了新思路。该算法通过时-空滤波去除运动矢量场噪声,根据运动注意力形成机理定义运动注意力模型;为提高注意力计算的准确性,定义了目标像素块的测度公式,采用D-S证据理论对多尺度空间运动注意力进行决策融合,最终获取运动目标区域位置。多个不同高清视频序列的测试结果表明,该文算法在全局运动场景中能准确对目标进行检测定位,从而有效克服了现有算法的局限性。

The detection to target in motion is a key technology in video analysis. This paper proposes a target detection algorithm based on a multi-scale motion attention analysis, which provides a new method for motion target detection under a global motion scene. Firstly, the noise of motion vector field is removed by filter, and according to the mechanism of visual attention, spatial-temporal motion attention model is built; then the trust degree of motion vector is suggested on the basis of validity analysis of motion vector, and decision fusion of multi-scale motion attention is accomplished by D-S theory for detecting the region of motion target. The test results of different videos show that the algorithm is able to detect precisely targets under a global motion scene, thus effectively overcoming the limitations of the traditional algorithms.